ULTIMATEEXPERIMENTSCommodore 64 · Work in progress

SuperEditor optimization and simplification plan

Updated 2026-10-27. This is the execution priority for the remaining M8 work. It preserves the original product scope and supersedes any proposal to keep adding optimization layers to the present renderer. It is a plan, not a claim that the budgets below have been achieved.

Current execution checkpoint

Commit a5af575 completed the first graph replacement: eleven semantic, query, lexical and rebase capsules plus their observer writers are gone. The runtime is now51 images and174,889 raw payload bytes, down41,699 bytes from the connected control, with unchanged resident/BSS/ZP/ReadyOS reservation. R33 is one checked current-BODY plan rather than another family of cache supervisors.

Commit ade8326 removes two measured sources of repeated work. C8 passes its already found Plain GOTO logical start to C23 through a full-key, CRC, VALID-last descriptor; the66K CPU path falls162,038,336→17,165,027 instructions and22,145→187 loads, with a real D81/FileOpen/IRQ PAL run at 2,344,684 cycles. R33 now keys the BODY plan on29 logical identity bytes, so transport/window moves do not force reparsing; A/B/A falls1,299,975→694,204 instructions. Neither change adds an image, bank, supervisor or resident state.

The a5af575 PAL profile exposed 66 C26→R33 event crossings and an11,285,120 cycle WRAPPED TOP, but that diagnosis must be read as the pre-ade8326 state. The29-byte logical key keeps the completed plan valid across the warm BOTTOM→TOP/A→B→A path. Exact current CPU drops to2,431,845 instructions and24 loads, so the proposed C26 batching becomes pure overhead: +33 instructions on warm TOP, +27–31K on edits/cold construction and +249 code bytes. Preserve its atomicity failure and corrected guard proof as rejected evidence; do not merge it. Exact current PAL confirms the composed fix: WRAPPED TOP is9,013,034 cycles/24 loads, down2,272,086 cycles (20.13%) and94 loads from the pre-key state. Disk or capsule ordering cannot account for that runtime change.

This checkpoint integrates the first bounded R21 inner-loop refactor. It carries absolute-y, strip-y and band through the nine raster rows, hoists the packed/native row stride into existing scratch, and collapses the two style-bit tests. It adds no image, state, call edge or DMA and keeps R21 at4,095 bytes. Exhaustive4,096-case component comparison is exact; public PAL falls13,060,967→12,854,325 cycles on LONG and9,013,044→8,822,966 on WRAPPED. R21 remains the largest exact cost. Continue to remove work inside its existing job lifetime; do not revive the rejected per-glyph REU atlas or add another checked transport.

Commit 4fc6f45 removes C20's generic measure object and keeps only its fixed BSW9 PETSCII/width/style path. The complete eight-byte font header is still checked for every glyph, including corruption between BEGIN and STEP. C20 shrinks4,094→3,934 bytes; exact current LONG falls1.823% and WRAPPED0.164%, with unchanged pixels, geometry, loads, calls and parser work. Commit d50545c then deletes30 R21 bytes that wrote VALID/baseline/height fields with no linked reader, saving2,390–3,180 instructions per tested public TOP. Both changes add no state, cache, image, call edge or supervisor.

Commit 1d78734 follows the same proof through C20 and deletes34 more bytes of VALID/reverse/index/width/baseline/height stores with no reader, saving11,187 instructions on LONG. An attempted active-hit loader shortcut is rejected even though PAL improved0.52%/0.32%: raw owner or epoch corruption while active no longer returned stale, so it weakened a checked invariant without a replacement.

The resident catalog now matches the exact51-image product rather than retaining thirteen unreachable rows. This deletes78 warm bytes and8,475 instructions across the complete51-image import sequence, with unchanged PRG/BSS/ZP, full16MiB REU, lifecycle and stack68 proofs. This closes the catalog-capacity cleanup. Further disk or catalog rearrangement is not an optimization priority: subsequent work must remove material foreground edit, parse, layout or raster work and qualify the whole public operation.

Commit 381c6a5 removes the repeated GEOS suffix-construction pipeline from ordinary Full edits. R28 now admits the edit once, C21 compares the complete 842-byte following row while it is resident, exact convergence terminates the suffix, and at most one signed-anchor rebase remains. The C21/R28 pair shrinks 7,406->7,288 linked bytes. The broad Full insert falls907,617->825,542 CPU instructions,383->310 DMA calls and168,061->166,782 DMA bytes, while all36 code loads remain. This is a material reduction in repeated work, but it does not meet the stock1MHz interaction target.

Exact-current PAL timing makes the remaining architectural deficit explicit. An ordinary Full prose insertion settles in825,542 CPU instructions and 3,116,645 PAL cycles. A heading insertion takes8,192,899 cycles, an inline strong insertion9,137,454, and a line break7,968,166. A styled bottom-edge GEOS DOWN costs2,559,990 instructions,338 code loads,868 DMA calls and 1,390,679 DMA bytes because21 authenticated visible rows are parsed and laid out again. C20, source staging, C24 semantic classification and C25 lexing alone repeat1,089,996 instructions on that scroll. These costs, rather than D81 file order or fixed REU slot placement, select the next work.

The first selected replacement was an early native8 retained one-row scroll transaction. It must rotate the existing22-row FE page, retain the overlapping bitmap and attributes with two checked DMAs, and run the existing parser/layout/painter for only the exposed row. It may add no image, REU bank, allocation, BSS, zero-page byte, resident state or cache family, and the total linked payload of every changed consumer may not grow. The promotion target is at most300,000 CPU instructions and four code loads for the exact public scroll, with byte-exact pixels, attributes, anchors, source and fallback behavior. If the transaction cannot preserve the outgoing FE/GCC authority before allocation, it is rejected rather than weakened or moved into another overlay.

The exact 381c6a5 retained-ring implementation fails that size gate before a runtime claim is possible. The strict authority bridge and ring infrastructure alone are28 linked bytes larger, overflowing C23 by19 bytes and C22 by13. An intentionally incomplete mechanical commit reaches+315 bytes, with C21 at4338 bytes,242 beyond its slot, while still omitting mandatory prefix authority, caret/semantic refusal, rollback and cancellation. The representation is therefore rejected. The2.56M-instruction scroll remains a measured deficit, but the next candidate must replace a larger existing C22/R21 preparation/raster route instead of adding ring machinery beside it. Evidence: build/retained-ring-scroll-probe-381c6a5/REPORT.md.

The current-head C22/R21 audit then establishes that the presumed larger route has already been replaced: C22 remains resident while the complete logical-row R21 job rasterizes and polls. Replacing its remaining CPU packet copy with a checked FE-to-row DMA fits without new state and shrinks C22 by3 bytes, but saves only0.30-0.43% on complete Plain/Markdown/Full and long Full page builds. It passes48 clip/style/selection frames and3,077 observed C22 calls, yet is rejected at the material gate. Further work must attack a different measured instruction class rather than repeat the already integrated row-residency design. Evidence: build/current-row-resident-prepare-381c6a5/REPORT.md.

The first owner-local checkpoint deletion is also sealed as a rejection. It fits as a50-image product, folds the cold font importer into AUX and removes 1,068 linked payload bytes. Plain GEOS insert improves234,701->184,177 instructions and3->1 loads. Full insert, however, regresses825,586->853,937 instructions despite36->34 loads because the stale post-publication key forces one C21 VERIFY plus17 R28 REBASE calls and18 C21 continuations. The added 31,454 R28 and683 C21 instructions exceed the deleted AUX/R14/resident work by 28,351. The global worker cannot be removed safely until checkpoint maintenance is owned at the actual C21/R28 publication boundary; merely moving it before publication loses the key needed by the next Full edit. Evidence: build/runtime-gap-audit-current/R30_C32_OWNING_PROBE_RESULT.md.

A second attempt moved the refresh into R28 at the apparent C21/R28 commit boundary. It deletes the AUX/R14 tail, saves583 linked bytes, and improves the measured Plain and Full inserts by52,221 and51,831 instructions with two fewer loads. It still fails the authority gate: the common Full fallback later publishes through C21/R19 and restores the old revision/length checkpoint 0/70 after the page has become1/71. R28 ADMIT and COMMIT placements therefore cannot own the final key. The candidate is rejected before broad gates or the optional font fold. Any successor must begin at the existing final R19 checkpoint publisher and pass repeated-near-edit plus fallback exact-key tests before performance work. Evidence: build/runtime-gap-audit-current/C21_R28_PUBLICATION_BOUNDARY_RESULT.md.

The managed resident overlay path now maps the sealed (slot,ID) pair directly to its dense51-entry catalog index instead of linearly scanning all imported images on every call. Import still publishes the exact index VALID-last, and every execution retains binding, provider owner/epoch, verified-backing CRC, active-slot and checked-DMA guards. The resident CODE segment shrinks2 bytes with no BSS, ZP, bank or image change. Clean Plain edits save1.84-2.02% of the complete operation; Full prose saves3.12-4.02%, and a forced-fresh diverse Full page saves8.58%. The claim is capped by the clean lookup-only comparison because some Full fixtures also take fewer calls and DMAs. Evidence: build/resident-overlay-call-current-381c6a5/REPORT.md.

The wide-mode presentation fault is also fixed without another framebuffer or hidden RAM. A mode command now publishes only the logical target. Grid60/80 select bitmap hardware after the successful final painter; a return to40 calls the existing ROM-font restore immediately before the first text-screen write. The measured40->60 all-zero interval falls from1,872,053 PAL cycles to zero and settle time remains neutral at7,232,112->7,231,485 cycles. The gridedit/text40 pair shrinks one byte and adds no state. This is deliberately described as a non-atomic presentation fix: the one shared display image is progressively overwritten for roughly0.29M-4.56M cycles depending on the transition. Exact final pixels, attributes, color low nibbles, faults, input cancellation, menus, GEOS smoke and IRQ/NMI bounds pass. Evidence: build/final-scope-audit-current/publication-deferral/QUALIFICATION.md. On the combined linked product the dense lookup also cuts68,096-114,493 PAL cycles from the measured mode-change routes and3,144,321 CPU instructions from the aggregate96-frame styled grid suite, without changing its overlay transport counts or publication order. Evidence: build/final-scope-audit-current/publication-deferral/COMBINED_DENSE_VALIDATION.md.

Product and machine constraints

Design for a stock 1 MHz C64 first. Keep PETSCII/SuperMark, all 40/60/80-column source views, GEOS-font source editing and formatted preview, session appearance settings, BASIC/assembler highlighting, folding, keyboard menus, one status line, file operations, selection, clipboard and undo/redo. Keep assembly micromodules, D81 development, logical REU allocation, the ReadyOS provider and suspend/resume boundary, and the untouched $C600–$C9FF reservation. JSON, UltimateDOS and actual ReadyOS application integration remain outside scope.

The assessment's proposed smaller Core/Deluxe editions are not adopted: they would remove requested features. Remove unnecessary implementation machinery, duplicate representations and unrequested extensions instead. No new visible features are to be added during this work. A source micromodule need not be a separately swapped runtime image.

The physical Ultimate is a user play session. Upload, mount, reset or load only after an explicit user request. Development verification stays in isolated VICE runs with warp/traps, plus CPU probes where useful. Measure emulated stock-speed cycles even when VICE runs faster than real time.

Inputs and verified diagnosis

Read both complete discussions: Assess Screen Refresh Optimization (01a17b30-b9e9-7f03-8ed2-5d53af23c291) and Assess Super Edit overlays and bloat (01a17b32-0554-7fa2-aa7f-10950c018c10). Their user messages and final assessments are retained locally in agentsworking/research/2026-10-27-optimization-thread-inputs.json. Recommendations are inputs, not automatic requirements or established facts.

The committed 12bde7a fixture was independently checked for this plan:

Item Baseline
Registered images 62:24 R-slot,38 C-slot
Linked payload, including embedded data and padding 220,651 bytes
Reserved fixed image slots 253,952 bytes in four REU banks
Images at least3,500 bytes / at least4,000 bytes 46 /30
Last16 images' payload 57,350 bytes
Available R/C execution windows Two4 KiB slots
General BSS / limit before ReadyOS reserve 1,534 /1,536 bytes

Payload is not synonymous with executable instructions. The inventory must separate instructions, tables/fonts, padding and duplicated routines. Source line counts including probes, generated code and test tools are not product code size. Host evidence storage is also a separate problem.

Measured problems already justify architectural work:

Use warm input attribution, the local ignored raster-attribution record at src/render/geos/raster-cost-probe/COST.md, the existing cache redesign, and the pinned geoWrite source study in agentsworking/research/GEOWRITE_REFRESH.md. Historical measurements are scoped baselines, not a new common-build benchmark or a measured geoWrite comparison.

The three existing parallel probes were stopped at bounded checkpoints for this plan. Their results change what gets built next:

Experiment Evidence and decision
Layout-only memo Whole CPU A→B→A saves2.94–10.91% in Markdown/Full but still rerasterizes. Do not add it as another permanent cache family.
Separate-bank page image Two images fit one optional64 KiB data bank, but the incomplete transfer draft is4,724 bytes:628 over its code slot. Never executed. Treat its storage census as input to a replacement, not a fit claim.
Corrected paint shadow mapping C21 needs4,139 bytes and R63 needs4,152:43/56 over their slots. Never executed. Replace delayed comparison and shadow storage rather than growing the slots.
Pre-shifted REU glyph atlas Synthetic20-row cold CPU work falls3,624,625→3,239,586 instructions, including construction. Complete verification per row costs7,328,001. There are999 checked per-glyph reads. Reject this as a production per-glyph transport layer.

Source/result notes are under src/render/geos/{layout-retention-probe, paint-shadow-probe,logical-row-job-probe/reu-atlas}/. The measurements are different scoped experiments; none establishes public native1 MHz latency.

Rules for choosing work

  1. Eliminate unnecessary work before making that work faster. Never rebuild a whole viewport merely because a local fast path refused an ordinary edit.
  2. Choose residency around an editing group, a logical row or a batch of rows. Parser, layout and paint helpers must not evict one another per character.
  3. Every production addition names the code, state and path it replaces, its deletion step, and the complete cost it removes. An isolated helper speedup is insufficient. No additional registered image by default; temporary migration images must have an explicit removal gate before release.
  4. Keep one clear fallback for each mode. An optimization miss must reach that correct bounded path without another chain of speculative caches. Preserve required representation differences between fixed and proportional text; do not invent one giant generic renderer merely to reduce file count.
  5. Preserve source, ownership, extent, generation, font and publication checks at real trust and yield boundaries. Hoist repeated checks only after proving a closed lifetime. Do not treat standalone REU ownership as proof against internal aliases, cancellation or stale derived state.
  6. Optional analysis and retained pages have bounded quotas and cheap eviction. They may save time; they may not make an ordinary edit fail for cache space.
  7. No new allocator, cache supervisor, edition split, setting or abstraction without a measured need. Keep the logical provider boundary. A replacement linear allocator is not justified by the evidence presently available.
  8. Total cycles and input latency decide. Image count, zero parser feeds, component percentages and smaller instruction counts are supporting metrics. Never add gains from different probes as though they compose automatically.

Proposed engineering budgets

These are initial targets to test at nominal1 MHz, not achieved results or new claims about hardware. PAL cycles and approximately equivalent milliseconds must be reported separately. Missed budgets remain visible; do not quietly increase them or conceal them with a narrower favorable case.

Interaction, settled current mode Cycle target Code-load target
Warm 40-column caret 15,000 0
Warm 60/80-column or GEOS caret 40,000 0
Ordinary 40-column same-row edit 30,000 0–2
Ordinary 60/80-column or GEOS same-row edit 100,000 0–2
Visible paragraph Return/BACK split or merge 250,000 At most 4
One-row viewport scroll with retained overlap 300,000 At most 4
Unchanged A→B→A return, if retained 150,000 At most 2
New diverse GEOS viewport, first coherent completion 500,000 At most 12

The original modest-document full-redraw goal remains approximately 0.5 s. A one-million-cycle new GEOS viewport is an intermediate improvement, not completion of that goal.

Input must be serviced at safe boundaries within 50,000 cycles of foreground work. A completed obsolete frame is not a prerequisite for applying queued input. Measure first visible feedback, latest-state completion and cancellation delay independently. The keyboard ring is finite; prove backpressure/overflow behavior under repeat and bursts rather than promising unlimited type-ahead.

The initial consolidation target is at most 48 registered images and 176 KiB linked payload, about 20% below the 62-image baseline. Phase 1 must check the actual section inventory before fixing the final release ceiling. Count and payload reductions cannot compensate for missed interaction budgets. More REU data can be sensible when it deletes CPU work and management code; count its construction, refresh, pressure and restore costs explicitly.

Ordered execution

0. Freeze a safe baseline and contain the experiments

Keep committed 12bde7a as the control. Record its exact sources, D81, fonts and settings. Preserve the uncommitted63 work and both structural regressions as evidence, without treating it as the accepted product. Stop opening additional probe variants until the present probes have a measured accept/reject result. Keep useful techniques from them, not their accumulated supervisor graph.

Before large code changes, commit a clean, reviewable checkpoint of the actual source on the dedicated optimization branch. The existing supereditor/reu-cache-redesign branch is suitable; do not sweep unrelated changes or artifacts into it. Ignored logs, generated binaries and external GEOS assets stay out of Git. Do not commit a known-broken experiment as a qualified release. Preserve it separately as a clearly labeled checkpoint if needed. No worktrees.

Exit: a reproducible control and a list of every retained or rejected probe, including known faults. No ambiguous mixture of committed and private claims.

1. Produce one common-build cost and residency inventory

Capture the same corpus and public operation sequence on the exact baseline: all four source views, preview, Plain/Markdown/Full, highlighting on/off, ordinary and styled text, BASIC/assembler, narrow glyphs at maximum row count, wrapping, selection, folds, and source positions above64 KiB.

Include first/steady typing, caret immediately after typing, CR/BACK at the viewport bottom with scrolling, line/page navigation, A→B→A, popup open/close, first/reopened preview, undo/redo and repeated edits through flush. Measure complete source processing, first pixel and completed publication, overlay loads/calls/DMA, IRQ/input queue, parsing/layout/glyph counts and exact source identity. Keep cold setup, warm work and idle analysis separate. CPU profiles select work; normal-speed emulated PAL cycles qualify it. NTSC is a final gate.

Generate a linked closure inventory and an ordered call graph per operation. Classify each image and shared state as hot, mode-specific, or cold. Identify duplicate source readers, copies, CRC loops, semantic stores and fallback controllers. Budget a new RAM/phase map with real lifetimes and IRQ access; recover space by deletion instead of squeezing more aliases into spare bytes.

Exit: one concise table containing the current deficits, chosen replacement boundaries, byte budgets and exact code to retire. This replaces selecting the next optimization by whichever overlay happens to have a few bytes left.

2. Keep the hot editing pair resident

First remove the measured C6→R30→AUX→R14 eviction cycle. Move the small post-publication freshness update into the owning transaction or defer it with an explicit invalidation/rebuild contract. Preserve source checkpoint freshness and undo boundaries. Consolidate source access and common projection work into the active edit/mode job. Keep cold dialogs/settings/preview/lifecycle in cold overlays. Do not move an entire large worker into the already-full resident.

Replace the executable$5000 staging arrangement with a budgeted residency map in which fonts/data cannot silently evict hot code. Source micromodules remain separate and link into operation-sized images. Removal of the stage stamp and re-prime paths follows only after the replacement covers their consumers and meets the complete-operation budget; a slower clean-looking map is not a win.

Exit: ordinary repeated keys meet0–2 loads; no forced post-key eviction; the replacement deletes obsolete routing/staging code. Local/caret gains must survive switching modes, opening menus and coming back to typing.

3. Replace fallback re-layout with retained lines and exact convergence

Use one bounded retained logical-line/block model containing source anchors, incoming/outgoing parser and inline state, font/style/width, geometry, baseline, fold projection and paint inputs. Reuse existing valid records where possible; do not add a second semantic truth beside type53 and then maintain both.

Restart at the affected logical boundary, including the previous line for CR deletion. Invalidate the affected complete inline BODY when a delimiter can change earlier styling. Propagate only until unchanged source and complete parser/font/style/geometry state converge. Fences can legitimately propagate to EOF; do that incrementally with cancellation, not in a blocking repaint. Retain an unaffected horizontal prefix where its pixels and attributes remain valid. Long lines require continuations rather than repeated scans from BOF.

Retain visible and adjacent lines under one bounded replacement policy. Move overlapping bitmap rows/attributes and regenerate only newly exposed rows on scroll. View/caret/selection changes must not restart parsing. Palette changes retain semantic/layout data; metric changes retain parsing but invalidate dependent geometry. Highlighting disabled must execute no lexer feeds; an unchanged block must not be lexed again merely because the viewport changed.

Compare each genuine newly completed row with its old packet at the row-ready boundary. A proven single842B old-row scratch plus accumulated dirty strips can replace 22 persistent shadows, their scattered addressing, the delayed comparison pass and duplicate publication certificates. This is a replacement design to prove, not an implemented fix: C20/source/parser leases must have ended before the scratch is borrowed, and source/font/provider/publication authority must survive the whole comparison and commit.

Replace the semantic/lexical rebase chain and its duplicated routing as the new model proves coverage. A prototype page memo that saves10% but leaves the dominant raster work and adds another full supervisor is not the endpoint. Store a prior page image only if it fits the same bounded policy, cheaply reuses genuine caret-free geometry and deletes work/code overall. Handle cache pressure as a miss, with no extra bank assumed to be available.

Exit: public ordinary/wrapped/CR/scroll operations stop at proven convergence; retained A→B→A skips both measurement and rasterization when its image is valid. Trace deletion of displaced semantic/lexical/paint paths, not just their bypass.

4. Replace strip-by-strip glyph preparation with a whole-row compositor

Compose each8/9px logical row once into its containing VIC bands. Bind the source/packet/font authority once per closed job, prepare glyph descriptors once, and carry the exact ink weights across shared8x8 color cells. Preserve bold, italic, underline, reverse, native fallback, selection/caret union, clipping and role priority/ties. Keep foreign font metrics correct.

Batch enough work while the compositor remains loaded to avoid a helper swap per glyph or band. Prefer ordinary lookup tables and fixed shift chains where they remove real work. The present raw-bit decode already skips invisible rows: do not claim a50% gain from eliminating invisible loop passes alone.

Judge the existing whole-row and REU-atlas probes on complete public rendering, including packet gathering, actual checked DMA, first-build/readback, styles, fault checks and every yield. A pre-shifted font cache that saves a synthetic steady job but doubles cost when revalidated per row is not ready to ship. Prefer a small per-font/per-job prepared representation to another permanent glyph cache manager. Reject variants whose setup or memory ownership cannot be amortized safely. Share source/preview primitives where their semantics actually agree; do not force formatted and literal layout into one state union.

Replace C22/R21 preparation and the eligible old raster path together. If the new compositor subsumes paint convergence, delete the private R63 supervisor and its shadow/certificate copies. If it does not, no permanent R63 addition is justified until the net image/bytes/load budgets and correctness gates pass.

Exit: one traversal per glyph per logical row, exact color-cell output, public cold/warm cycle wins, and retired old preparation/staging paths. Do not preserve every historical shortcut as another fallback beneath the new implementation.

5. Make responsiveness independent of renderer throughput

At safe yield boundaries, detect queued editing input, stop obsolete paint, apply a bounded batch and merge damage against the newest revision. Preserve command ordering, modifiers, physical key identity and undo grouping. A held key must neither starve all publication nor force every intermediate state to paint. Do not roll back an accepted edit when an optional derived job stops.

Retain the existing visible publication until a correctly sourced replacement is ready; stage and commit bounded dirty regions using a proven memory map. Logical publication atomicity is not a claim of tear-free VIC frame timing. Menus restore their covered rectangle; help/status changes do not reparse the document. Test typing during worst-case reflow and scrolling at stock speed, including buffer saturation, not just injected keys after every idle state.

Exit: bounded input service/cancellation latency, exact accepted text, no starvation, no stale result published after a newer edit, and no unnecessary full-page redraw on popup dismissal.

6. Cut startup and distribution baggage

Separate a lean runtime D81 from the existing diagnostic/fault-fixture D81. Remove test PRGs, malformed overlays and probe assets from the runtime disk; retain them in the development package. Keep all requested runtime features.

Measure startup by disk open/read, CRC/import, code installation and font installation. Bring up40-column editing without eagerly loading proportional fonts, preview workers, lifecycle copy code and every other cold feature. Evaluate a simple sequential indexed overlay pack against many sidecar opens, using actual KERNAL/D81 capabilities; do not assume random seek or UltimateDOS. Retain per-image bounds/identity checks and meaningful missing-asset errors. Warm REU swaps remain distinct from disk first-use loads.

Exit: no unrelated eager assets before first useful edit, at least80% less time-to-ready than the measured approximately11m32 hardware baseline on the same configuration, and separate measured first-use times. This is an initial reduction target, not acceptance of a multi-minute editor startup. First-use costs must be disclosed rather than hidden after the initial screen.

7. Delete, qualify and publish one result

After each replacement, delete its retired production images, state, build paths and dispatch branches. Keep historical probes outside the runtime closure. Use a short before/after table: bytes, images, actual loads and code DMA, cold/steady cycles, input delay, memory and remaining failure cases. If a candidate adds machinery without meeting a complete-operation budget, reject it or redesign the boundary; do not promote it because qualification took work.

Require one common final source/settings/font cohort for all-mode editing, scroll/CR/selection/folds/menus/open-save/undo-redo/preview, PAL and NTSC, lower REU pressure and16 MiB, and the existing lifecycle seam. Keep independent pixels, exact source/history/clipboard, ReadyOS reserve, typed memory, ownership and fault/cancellation checks. Preserve the two new structural regressions. Update maps and stack analysis after actual call-graph changes.

Run focused CPU/probe tests while choosing a design; run the affected native matrix when a candidate is credible; run the complete final cohort after integration. Do not recapture unchanged broad galleries or run huge evidence seals for every small rejected experiment. Retain representative raw evidence, source identities and failure records without duplicating large artifacts.

Parallel ownership and immediate next action

Use at most three concurrent agents in this repository, without worktrees. Root owns production integration, shared builds, Git and native launches. Independent lanes own only their private sources/results: residency/code census; retained-line/scroll design; compositor measurements. Each closes with one accept/reject recommendation and a deletion list. Serialize overlapping memory-map decisions. Do not message the two assessment threads or deploy to hardware as part of adopting this plan.

Immediate next action: finish the bounded current probes, preserve their findings, and produce Phase 1's common-build cost/residency table and replacement map. Then tackle the hot-key eviction loop and retained-line fallback. Do not continue filling the64-entry catalog with independent optimization supervisors.

Current ruthless deletion pass

The catalog-capacity change is closed. Current production work must delete live instructions, state, repeated validation or raster work and report the effect on the complete foreground operation. Disk order and catalog layout do not qualify as performance work in this pass.

The first resident deletion removes the 28-byte clipboard compatibility dispatcher, routes its sole live COPY STEP caller through the controller's existing C10 dispatcher, and deletes eight unread state bytes. The linked result is 26 fewer CODE bytes and eight fewer resident state bytes. A successful copy step saves 15 CPU cycles and one resident call; overlay calls and loads stay unchanged. This is a code-graph simplification, not a redraw claim.

The next two independently measured candidates target foreground work directly: C32 typing removes repeated guards and validation wrappers, while R21 raster specializes its production-only full-width contract and deletes generic clipping. They must retain exact pixels, source and fault behavior after composition before promotion. Continue from whole-operation PAL attribution after those cuts; do not return to overlay-file rearrangement.

The C32 cut qualifies independently: it removes 26 staged bytes, saves 2,111 instructions on steady INSERT and 2,008 on steady BACK, and improves native PAL steady operations by 5.53% and 5.64%. The 46-frame public matrix, eight API fault cases, exact pixels/source/REU checks and the composed 51-capsule package pass. Caret work and overlay loads/calls remain unchanged.

The R21 cut also qualifies: C22's sole row setup always supplies LEFT=0 and RIGHT=320, so R21 now rejects any other pair and deletes the unused partial-width implementation. It removes 191 executable bytes and 1,200 pairs of horizontal resolver calls per full TOP operation. Public CPU work falls 2.34–3.73%; native PAL LONG and WRAPPED fall 1.68% and 2.40%. Exact pixels, attributes, source, fault-before-write behavior, stack, overlay traffic and DMA remain unchanged. The role-zero shortcut remains rejected because it regressed Markdown and Full.

The next R21 pass hoists band-invariant role initialization and VIC attribute address calculation out of the 1,200-tile resolver. It removes 27,690 R21 instructions from LONG and WRAPPED, saving 99,052 and 99,398 native PAL cycles. It is one executable byte smaller, adds no state or calls, and keeps component, fault, stack, package, pixel, load and DMA checks exact. The histogram-pointer variant remains rejected because it added code for a weaker resume contract.

C32 now writes each caret-free edited row directly into the existing two-band cache-block layout and reuses that copier for VIC publication. This deletes the separate 9x40 scatter pass and shrinks TYPE by 37 bytes. Complete steady edits remove 1,845 instructions and improve native PAL INSERT/BACK by 4.14–5.03% across Plain, Markdown, Full and styled text. No state, stack, overlay or steady DMA work is added; the full edit/fault and native pixel matrices pass.

Resident input routing now releases style ownership once before dispatch and bypasses the public command cleanup wrapper on internal routes. Direct command callers keep the wrapper. This removes 25–28 cycles and one call frame from each accepted input command and deletes two resident CODE bytes. It is a small common path cleanup alongside the material edit/raster cuts; it does not change overlay or REU traffic and is not counted as a redraw solution.

R21 glyph measure/decode now consumes the fixed BSW9 contract already certified at job admission. It deletes per-glyph style, pointer, width, start/end and vertical-visibility revalidation plus two private pointer helpers. This removes 189 executable bytes and 9,647/9,240 helper calls, reducing native PAL LONG and WRAPPED by 6.10% and 8.45%. Job-level validation, malformed-job refusal, pixels, source, stack, overlays and DMA remain exact; no table, cache or REU path is added.

Cold state initialization now relies on its linker-bounded complete BSS clear instead of zeroing six render bytes again, and it reuses the loop's proven X=0. Zero-byte linker assertions keep those fields inside BSS. This removes 16 live COLD bytes and 51 startup cycles with identical dirty-RAM results. It is recorded as bloat cleanup only and carries no foreground performance claim.

Material replacement boundary after the bounded cuts

The bounded instruction deletions above are complete. Further leaf-loop and overlay-routing variants are closed unless a new whole-operation profile shows them to be material. Three exact-current experiments establish that boundary:

An R33/C26 direct span accumulator removed 66 event crossings and reduced a forced WRAPPED insertion by 149,077 instructions, but that was only 4.13% of the complete operation and grew the two images by 408 bytes. It is rejected. This is the decisive lesson for the next work: crossing, load and DMA counts are diagnostics; only deleted live bytes and complete foreground PAL cycles qualify a production change.

Implemented Plain direct-row slice

The first replacement is intentionally smaller than the proposed retained LRV3 object. The proposed persistent 96-byte header, 24-record directory, style canvas, parser carries and manifests had no next-edit reader: every Plain edit would have rebuilt and written them before consuming the same information once. That storage and its REU DMA/CRC work are deleted from the live design. Internal lrv3 source names remain temporarily, but they describe a local transaction, not a persistent retained-line cache.

C6 captures a one-edit certificate from the last completely published resident layout before mutation. After mutation it supplies only a physical-row hint. R25 frame 18 rechecks the certificate, Plain capabilities, width, post-edit length and revision, document/REU epochs, and the hinted se_v_rows start and following anchor. It derives that row's CR-exclusive length, proves its CR/EOF terminator and the edit position, and accepts only a row that remains shorter than the active 40/60/80-column width. The selected start, length, flags, row and caret inputs are stored in the existing $8D00 phase union; PREPARED is written last. There is no syntax parse, styled carry, directory scan, REU LRV publication or manifest authentication on this Plain path.

C33 frame 1 enters the direct painter only for a current prepared transaction; otherwise it preserves the ordinary warm-caret path. The painter rebuilds and publishes one row, commits the existing native row packet for 60/80 columns, shifts later resident row anchors and the bottom anchor by one byte, and retires the transaction. CR edits, wrap/row-count changes, selection, folds, styled content, stale witnesses and faults take the established cold fallback. This is a deliberately narrow first slice; Markdown, Full and GEOS retain their current correct renderers until a later replacement can delete their live work.

The obsolete fixed-grid graph is gone, including SP2, its C14/R14 dispatch and reader glue, the R14 paragraph shifter, and the displaced C33/R25 type/long families. Current linked sizes are resident main.o 271 bytes (three bytes smaller), C6 3,570, R25 3,231 and C33 3,784 bytes before its fixed native-font segment, leaving 56 bytes before $BF00. Against the exact 5,220-byte retired graph, additions are C33 1,348, R25 1,106 including 12 bytes of RODATA, and C6 343; the resident saves three bytes. Replacement code is 2,794 bytes and the net deletion is 2,426 bytes. The change adds no BSS or zero-page state and no image, bank or catalog entry. Do not revive a persistent directory or styled publisher until it replaces a measured live consumer and remains a net deletion.

Static actual-6510 probes pass 23 certificate cases and 12 row-hint boundary cases; a separate check proves stale PREPARED transactions are retired. The actual-product Peddle gate also passes exact source, pixels, attributes, row anchors and caret: insert/delete costs are 10,548/10,243 instructions at 40 columns, 87,619/87,345 at 60, and 50,689/50,460 at 80. Those ordinary Plain operations meet the 30K narrow and 100K wide instruction targets. They use 3/3, 6/6 and 6/6 overlay loads and 6/6, 8/8 and 8/8 calls respectively, so the wide-mode load target remains open. Warm caret is 5,179/33,698/34,859 instructions. Correct cold fallback costs 317,297 for an 80-column CR insertion and 426,997 for the 60-column wrap-boundary refusal.

Native PAL VICE now confirms the wide Plain path with exact source and reviewed screenshots. At 60 columns insert/delete publish first at 364,043/363,415 cycles and settle at 364,191/363,520; at 80 they publish at 218,387/217,945 and settle at 218,492/218,050. Each operation publishes one draw sequence and uses six loads, eight calls, nine DMAs and 23,072 DMA bytes, with only the expected grid60/grid80, R25, C6 and C33 groups. The tracked native scenario now explicitly issues syntax plain after opening the fixture; the prior setup did not prove a Plain route. These spans run from public input entry to foreground quiescence; they are not physical keyboard latency. NTSC and physical-hardware timing remain pending. Overlay counts remain diagnostic; rearranging disk overlays earns no optimization credit.

The original product scope remains unchanged. After the Plain slice finishes native qualification, the next work is whichever measured styled fixed-grid or GEOS parse/layout/raster boundary can delete the most live work without adding another parallel authority.

Styled wide-grid single-publication checkpoint

Full and Markdown 60/80-column edits previously published an unstyled bitmap, then completed the one full style pass and published the same viewport again. The wide C16 producer now returns ST_PENDING after checked leaf admission and style-context binding but before any cell or cache write. Resident clears the strict draw-busy flag and returns to the normal idle loop, so queued input is checked before C14 receives its first bounded slice. Once the style plan is valid, C16 paints and finalizes once. Nonzero UI bypasses withholding, and an exact-context planner failure schedules one plain fallback redraw. The unsafe 40-column variant was removed because R1 produces wrapped anchors during paint; prepaint styling there bound a valid plan to stale anchors.

This is a scheduling deletion, not parser convergence. Full 60-column work falls from 2,649,645 to 1,000,409 instructions (62.24%); Full 80 falls from 1,805,177 to 1,010,676 (44.01%). Loads fall 29 to 27 and DMA falls from 126 calls/135,025 bytes to 76/116,181. C16 GRID_FINALIZE falls from two to one. The linked ledger is exactly neutral: resident +14 bytes, text40 -15 and gridedit +1; no BSS, zero-page, REU, image or catalog growth. Full pixels/attributes, Plain 40/60/80, structural CR, UI/input interleaving, D81 and ReadyOS-reservation gates pass.

Normal-speed PAL VICE confirms the whole-operation reduction. Full60 settles at 3,996,868 cycles instead of 9,897,057 (59.61% lower); Full80 settles at 4,043,622 instead of 6,813,027 (40.65% lower). Both preserve independent source/bitmap/attribute, catalog, stack and REU oracles and publish exactly once. Loads fall 29 to 27 and PAL DMA falls from 123 calls/134,123 bytes to 75/115,759. The remaining roughly 310K-cycle tail is the existing post-frame C33 cache capture with nested R25 authentication, not syntax parsing; retaining that next-input authority is a separate replacement decision, not an overlay placement exercise. Evidence: agentsworking/evidence/M8/20261027T101239594697Z-display-timing-full60-80-normal-pal-post-change/normal-speed-full60-80-result.json.

The remaining styled boundary is still material: C14/C15 continue a complete visible style pass before that single publication. Their current 7,864 linked bytes contain 4,837 bytes of parser cores and 3,027 bytes of replaceable orchestration. The next replacement must reuse the dormant checkpoint region for authenticated incoming/outgoing logical-line carries, stop at exact parser state convergence, skip lexical work for unchanged code lines, and delete the displaced orchestration in the same change. Overlay placement alone earns no credit.

Styled row transaction: implemented replacement

The first convergence prototype proves the useful part of the idea and also defines what must be deleted. On an admitted Markdown 80-column edit it stops the block and inline parsers at the changed logical-line boundary: 24 and 23 feeds for the representative prose insertion, versus 129 and 52 on the structural CR fallback. The complete CPU operation is 337,407 instructions, versus 803,220 for that fallback, with exact source, pixels, attributes, anchors, caret, undo and redo. The prototype is not the production endpoint. styler plus viewsource grows from 6,373 to 7,456 linked bytes, an exact 1,083-byte increase, and the accepted path still runs the old C11 layout and post-publication C33/R25 capture. Moving parser code between disk images is size-neutral and receives no optimization credit.

The retained no-preclear cut is independently proven. Sixty poisoned cold jobs cover Full/Markdown, all three widths, empty/EOF/CR/wrapped/BASIC/assembler documents and two poison values. No unwritten plan byte is read. Removing the grid fill loop and its counter saves 63 C14 bytes and up to 1,760 stores per build. The former two-byte counter is currently inert reserved padding so sealed parser/source addresses do not move; linked BSS is unchanged. This deletion survives regardless of which styled transaction replaces the prototype.

The production boundary now admits the one-byte edit directly in C6, seals a compact ticket and packed row carries in the fixed $4770 checkpoint window, and never enters R25 frame18 or C11 on an accepted edit. C14/R14 restores the incoming block state, performs bounded parse and planning, and marks the ticket ready. C33 frame1 authenticates that result, builds and paints the affected row transaction, calls the existing R25 native painter where required, rebases the suffix anchors and publishes exactly once. No accepted path runs C16 full-page render, C11 layout, frame15, post-publication capture or other work after the commit. This deletes work from the foreground edit graph; disk placement is unchanged and receives no performance credit.

The final checkpoint uses five-byte packed carry records and a CRC16 over all 110 carry bytes. The CRC occupies two bytes that the plan originally reserved as slack, leaving seven spare checkpoint bytes. This is an intentional change to the byte gate: arbitrary carry corruption is now detected before admission, with no new BSS, zero-page, REU object, image, bank or supervisor. C33 also applies native roles while it constructs valid source-cell packets. The former separate 40/60/80-cell style pass is gone, and padding cells receive role zero. Cold style construction no longer clears 880/1320/1760 plan cells or 566 state bytes that are proven overwritten.

Structural convergence refusal no longer reparses against pre-edit anchors. It retires the ticket and source lease, returns ST_RETRY, lets ordinary layout publish the new geometry, and only then starts a fresh cold style build. This ordering fixed the real 40-column :::region row-growth witness while deleting 31 C14 bytes. Wrapped row growth is also rejected before mutation when the new source length exceeds the ticket's certified row geometry.

The linked result meets the replacement byte gate: styler.bin is 3,311 bytes and viewsource.bin is 3,283, for 6,594 of the 6,630-byte combined allowance. window.bin is 4,071/4,096, grid.bin is 4,096 and geostail.bin is 3,965; resident state remains 1,533 bytes and the change adds zero BSS and zero-page bytes. The complete matrix passes all 72 cases. Its 48 admitted cases preserve source, pixels, colors, anchors, caret and history, publish once, use 6–10 loads and execute 24,065–129,441 instructions. Twenty-four structural cases, five stale/corrupt carry cases and queued-edit/F6 cancellation all pass. Registered stack use is 52 foreground plus 16 IRQ/NMI bytes, or 68 total. Evidence is in build/styled-grid-carry-matrix/20261027T121902469141Z/result.json and build/editor-registered-stack/20261027T1219-current51/proof/qualified-result.json. The exact final styled transaction still needs normal-speed PAL qualification; no PAL claim is inferred from the instruction-count harness.

The next measured cuts stay inside this transaction instead of rearranging its overlays. Replace per-cell 24-bit range, selection and caret comparisons with row-relative counters (estimated 2.5–4.5K instructions per 80-column row). After proving uninterrupted ownership, remove repeated 59-byte guards between serialized fast slices (2–5K/edit), eliminate empty C14 phase-transition yields (1–3K plus fewer calls), and stop copying then generically rebuilding the 256-byte cache header (about 2.8–3.5K). A sequential suffix-anchor walk can save another 300–500 instructions. The largest, riskier step is a native dirty-span row renderer, which could remove 30–70% of leaf paint work. CRC removal is not a candidate: it is now the tested corruption policy for persistent semantic state. The detailed implementation and acceptance record is maintained in agentsworking/STYLED_ROW_TRANSACTION_PLAN.md.

Row-relative packet construction and completion-guard deletion

The first post-transaction cut removes work inside the accepted foreground path. C33 now computes the 24-bit row offset, remaining content count, endpoint state and caret column once per row. Its cell loop no longer repeats 24-bit range, selection and caret comparisons for every one of 40, 60 or 80 cells. It clears character, role and selector bytes as it produces each owned cell, so padding cannot retain stale reverse or selection flags. Selection is rejected at the authenticated R25 admission boundary and is no longer retested inside the admitted packet loop.

The packet builder itself falls from 2,634 to 1,347 instructions for the ordinary 40-column case, 3,275 to 1,668 at 60 columns and 3,915 to 1,988 at 80 columns. Near-full rows fall from 3,266/4,897/6,527 to 1,639/2,420/3,200. The complete styled consumer now uses 7,498/7,306 instructions at 40 columns, 84,030/83,854 at 60 and 55,295/55,147 at 80 for insert/delete, with the same two or three loads. This is a ten-byte C33 code increase over the required packet-flag clear baseline; it adds no state, zero-page, image, bank or catalog entry and leaves 17 bytes before the fixed $BF00 boundary.

The focused production oracle covers an empty line, first/live/end/padding caret positions, width-minus-one content, exact-width CR/EOF refusal, wrapped and panned rows, a retained row beginning at source offset 66,000, poisoned character/role/selector planes, and selection refusal with the full 16 MiB REU unchanged. Evidence is in build/lrv-row-relative/20261027T124658181333Z/result.json. The complete 72-case styled matrix, including five corruption classes, queued input, F6 cancellation and stack 68, passes at build/styled-grid-carry-matrix/20261027T124651196416Z/result.json.

C14 also removes a second source guard at complete:. Every path to that label remains inside the same uninterrupted se_style_step that already authenticated the source, so the second R14 call could not observe a new source authority. The deletion removes eight linked bytes and 945–946 instructions plus one R14 call from every completed planner, without changing overlay loads. Structural wrap routes that complete two planners remove 1,892 instructions. The larger private guard/yield fusion experiment is deliberately not shipped: although it saved more instructions, it added 23 linked bytes and slightly lengthened the measured input-response bound.

The next admitted change must meet the same standard. A direct authenticated header update is under proof because it can delete two 256-byte CPU copies and the generic header rebuild while retaining both CRCs and valid-last publication. Dirty-span painting and dead graph removal remain candidates only if their complete-operation savings exceed their code cost. Moving the same work between disk members or overlays remains outside the optimization ledger.

Wide raster, header and dead-graph cuts

The next three cuts remove more work from the same transaction. The native 60/80-column leaf already clears all 320 bitmap bytes and 40 attributes before glyph painting. C33 now passes the authenticated live-cell count rather than the physical column width to the glyph loop, so blank padding is not converted and painted a second time. A poison oracle fills the entire old row and its attributes before insert and delete; exact output proves the unconditional clear still erases a longer old suffix. This costs three C33 bytes and removes 30,023–32,409 instructions at 60 columns and 17,493–18,157 at 80. The 40-column path is unchanged.

The wide NC header is now updated in place after its existing authentication. The transaction writes changed row CRCs, suffix anchors, bottom, caret, revision, source/style lengths and publication fields directly into the live header, then retains the existing final header CRC, guarded DMA store and valid-last publication. It no longer copies 256 bytes to $8500, copies them back, and calls the generic full-header rebuild. Twelve styled transactions cover 60/80 columns, rows zero through two and insert/delete; all 256 final header bytes are byte-identical to the control, and independent checks recompute the header CRC and every changed packet CRC. Header corruption, forged magic with a recomputed CRC, and physical-owner/DMA faults refuse without publishing.

An added Plain bottom-row oracle caught an inactive-descriptor mismatch that the styled cases could not expose: unconditional revision and length writes changed the inactive style key when FIRST+ROWS==22. The final code gates those style stores on the styled transaction bit while retaining unconditional source-key updates. All four 60/80 insert/delete bottom-boundary cases then match the full 256-byte control header. This failure is why a byte-exact final-object oracle is required in addition to screen equivalence.

Finally, four record-flag stores and their two reverse consumers are deleted: no current producer can set LRRF_REVERSE. Two se_selection_range calls are also gone because R25 already refuses marked selection before admission and C33 no longer reads the computed range. This removes 49 linked executable bytes and 147–193 instructions from each admitted edit without changing state, calls, DMA or resources.

With all three cuts composed, the focused styled consumer uses 7,327/7,141 instructions at 40 columns, 49,754/47,198 at 60 and 33,529/32,723 at 80 for insert/delete. Relative to the row-relative checkpoint, that removes another 171/165, 34,276/36,656 and 21,766/22,424 instructions. Plain wide insert/delete falls to 46,353/44,111 at 60 and 30,783/30,065 at 80; warm caret routing is unchanged. geostail remains below its 4 KiB limit and the fixed C33 image retains 41 contiguous bytes before $BF00; there is no BSS, zero-page, resident, image, bank or catalog growth. Exact-current full matrix, stack and normal-speed PAL qualification are the final gates for this composed slice.

Those gates now pass on D81 28e6041abecc16025e90c50d8cc525cac521b472b57dfed6923cf0c195ba185a. The fresh 72-case matrix contains 48 admitted and 24 structural cases with zero failures, all five stale/corrupt classes and both cancellation cases pass, and the registered bound remains 52 foreground plus 16 IRQ/NMI bytes. Admitted operations use 6–10 loads and 22,065–91,396 instructions. Evidence is in build/styled-grid-carry-matrix/20261027T130013415847Z/result.json and build/editor-registered-stack/20261027T125955250755Z/qualified-result.json.

Normal-speed PAL VICE passes Full/Markdown, insert/delete and all three widths with exact source, pixels, colors, anchors, history, catalog, payload, stack and shim witnesses. Every case publishes once and executes no parser or overlay call after publication. Settled 40-column time is 201,893–203,665 cycles, 60-column time is 394,267–400,529, and 80-column time is 317,301–320,254. This is about 5% faster at 40, 23.0–23.9% at 60 and 20.9–21.3% at 80 than the preceding qualified transaction. The exact native artifact is agentsworking/evidence/M8/20261027T125939500129Z-display-timing-styled-grid-convergence-baseline/normal-speed-pal-styled-transaction.json.

The matching normal-speed NTSC cohort also passes all 12 cases. Settled 40-column time is 203,915–205,245 cycles, 60-column time is 397,662–403,870, and 80-column time is 320,381–323,205. Publication is followed by a constant 104-cycle quiescence tail with no parser or overlay work. The complete oracle and screenshots are in agentsworking/evidence/M8/20261027T130055318500Z-display-timing-styled-grid-convergence-baseline/normal-speed-ntsc-styled-transaction.json and its sibling ntsc-styled-transaction-contact-sheet.png.

Deletion-only transaction cleanup

The next pass removes state and branches whose values are invariant in the live row transaction. LT_CLIP was always zero, so its stores, 16-bit additions and carry refusal are gone. Success no longer normalizes unused A/carry values or rewrites se_st_cp_valid and se_st_request to values already required by admission. The styled descriptor update is fused with the authenticated wide header loops; 40-column mode retains its required local copy because it bypasses that header path. The orphan LRV3 magic and geometry constants are also gone.

This deletes 69 linked bytes: 30 executable bytes from grid, plus 30 executable and nine read-only bytes from geostail. No BSS, zero-page, resident state, image, bank, catalog entry, load or DMA is added. Focused insert/delete work falls from 7,327/7,141 to 7,309/7,123 instructions at 40 columns, from 49,754/47,198 to 49,711/47,155 at 60, and from 33,529/32,723 to 33,486/32,680 at 80. This is a small hot-path saving coupled to a real code deletion, with no credit assigned to packaging or overlay placement.

The exact-current D81 is b6ceb3089774a55c7be5b9d0396341e6fde0e71b303b32aeabdd8a860af03552. The 72-case matrix, all five corruption classes and both cancellation cases pass at build/styled-grid-carry-matrix/20261027T132718007905Z/result.json. The byte-exact 256-byte header/CRC oracle passes at build/lrv-header-direct/20261027T132700114530Z/result.json, and the focused consumer and row-relative gates pass at build/styled-ticket-consumer/20261027T132708785100Z/result.json and build/lrv-row-relative/20261027T132710997907Z/result.json. The established 68-byte stack bound remains conservative because this pass only deletes instructions and checked branches from existing leaves and introduces no call, push, interrupt or recursion edge.

Strict 60-column renderer

The 60-column leaf no longer carries the generic 80-column branches and three per-scanline helper calls. Its strict renderer selects the fragment shift table once, runs one fused eight-scanline loop, restores the patched operand before return, and refuses every non-60 mode. It retains the existing 68-byte caller frame, four borrowed zero-page bytes, full-row bitmap and attribute clears, fixed font addresses, color arbitration, inverse, selection and caret behavior.

R17 executable code shrinks from 1,266 to 1,205 bytes, freeing 61 bytes before the fixed $A800 font. The standalone row differential passes 1,117 valid cases, 17 shared invalid frames, 32 random fonts, full output poisoning and a whole-RAM guard. The integrated styled consumer falls from 49,711/47,155 to 47,092/44,796 instructions for 60-column insert/delete, a 5.27%/5.00% cut. The full 72-case matrix passes at build/styled-grid-carry-matrix/20261027T132919561136Z/result.json; structural 60-column fallbacks also remove about 124,000 instructions. Modes 40 and 80 are unchanged.

Normal-speed PAL 60-column insert/backspace now settles in 378,794/373,471 cycles for Full and 378,895/373,495 for Markdown. This is 5.27–5.41% faster than the preceding qualified product. NTSC settles in 381,917/376,750 and 382,157/376,445 cycles, 5.28–5.43% faster. Both 12-case cohorts preserve exact source, pixels, colors, anchors, history, catalog, payload, stack and shim, publish once, and perform no parser or overlay work after publication. Evidence and screenshots are in agentsworking/evidence/M8/20261027T133119318729Z-display-timing-styled-grid-convergence-baseline and agentsworking/evidence/M8/20261027T133313781696Z-display-timing-styled-grid-convergence-baseline. The NTSC verifier now selects the native (32,23) 320x200 viewport for its 384x247 capture instead of assuming PAL's (32,35) viewport.

Fuse row-transaction state updates

The direct-row admission path now tests its existing retained-view marker before calling se_view_init; only callers that already proved and initialized that view take the bypass. The suffix-anchor loop also copies authenticated wide header rows while it applies the signed 24-bit edit delta. This removes a second walk, pointer setup and helper call instead of moving either operation elsewhere.

The change deletes four linked executable bytes and adds no state, call, load, DMA, image, bank or catalog entry. Focused insert/delete falls to 7,011/6,825 instructions at 40 columns, 46,360/44,064 at 60 and 32,754/31,948 at 80. That removes 298 instructions at 40 and 732 at 60/80 from the preceding strict-row checkpoint. The exact-current D81 is a087050d3b8c1e02d513e4c2ae806dbf7dddbf8dd798954b3b897de540d87df2; the 72-case matrix, corruption and cancellation gates pass at build/styled-grid-carry-matrix/20261027T133521703834Z/result.json, with the byte-exact header and row-relative gates at build/lrv-header-direct/20261027T133509620060Z/result.json and build/lrv-row-relative/20261027T133518223753Z/result.json.

Normal-speed PAL settles in 200,984–202,405 cycles at 40 columns, 371,372–376,571 at 60 and 314,955–317,806 at 80. Normal-speed NTSC settles in 202,949–204,289, 374,369–379,714 and 317,679–320,531 cycles respectively. This is a further 0.41–0.81% complete-operation reduction over the strict 60-column checkpoint, with unchanged load/call/DMA counts, one publication and no parser or overlay work after publication. Both 12-case cohorts preserve the full exact oracle and screenshots at agentsworking/evidence/M8/20261027T133548485400Z-display-timing-styled-grid-convergence-baseline and agentsworking/evidence/M8/20261027T133701134151Z-display-timing-styled-grid-convergence-baseline.

Delete empty parser phase yields

C14 and C15 now continue from parser initialization into scan/paint within the same existing bounded budget. A parsed byte and a painted cell both consume that budget; no slice becomes unbounded. R14 also omits its second source guard only for the private block step that C14 already guarded without an intervening yield. Public R14 frame 2 performs the guard before entering the same read leaf, so its authenticated API remains unchanged. The overlay dispatcher uses a shorter decrement chain to fund that public wrapper.

This removes six linked executable bytes and no state, call frame, load, DMA, image, bank or catalog entry. The exact-current 72-case matrix passes at build/styled-grid-carry-matrix/20261027T134952529642Z/result.json. Admitted edits save 2,669–10,378 instructions, with a 4,012 median; unchanged-code cases save 4,011–4,013 at every width. Structural fallbacks save 9,048–41,383. Parser calls/feeds and final syntax output remain exact. All five corruption classes, queued edit and F6 cancellation pass; input injected at C14 reaches resident service within 5,781 instructions, and the registered stack bound remains 68 bytes. The broader style-editor suite also passes, including public reader faults and asset removal at a real 32-unit lexical paint boundary.

Normal-speed PAL complete edits improve another 7.33–13.84% over the preceding transaction checkpoint: 173,158–174,678 cycles at 40 columns, 343,108–348,978 at 60 and 286,879–290,063 at 80. NTSC improves 7.38–14.11%: 174,323–175,736, 345,732–351,699 and 289,411–292,011 cycles. Each edit makes eight fewer overlay calls while retaining the same load count; the six-byte smaller R14 payload also removes six transferred DMA bytes. All 24 native cases preserve exact source, pixels, colors, anchors, history, catalog, payload, stack and shim, publish once, and do no parser or overlay work afterward. Evidence and screenshots are in agentsworking/evidence/M8/20261027T135637948567Z-display-timing-styled-grid-convergence-baseline and agentsworking/evidence/M8/20261027T135726493122Z-display-timing-styled-grid-convergence-baseline. The exact D81 is 47859cdb2579ecf983c1a6538e7b127d2a5460cebab50e6b6eb2d4761e49c3d0.

Direct wide status and anchor arithmetic

The retained 60/80-column row path no longer rebuilds the complete 14-cell status line through the generic TUI formatter. It compares the authenticated old and new caret nibbles, redraws only changed hex cells, and writes the fixed Modified label directly from the sealed compact native font. The status font is ordered for direct hex indexing, so its 26-byte screen-code lookup table is gone. Warm wide caret motion uses the same bounded delta renderer. The suffix anchor/header loop also sign-extends the one-byte edit delta once instead of recomputing it for every 24-bit anchor byte.

The optimization itself deletes 28 meaningful linked bytes: two from CODE and 26 from TAIL, with no BSS, zero-page, resident, image, bank or catalog growth. The separately committed $8D00 alias correction adds 11 linked bytes, so the current product is still 17 linked bytes smaller than the parser-phase checkpoint while restoring insert/Undo correctness. The alias fix clears the authoritative se_i_ready gate before either row-transaction producer first writes $8D00; C8 then rebuilds record zero before any checkpoint read.

The exact-current focused consumer passes at build/styled-ticket-consumer/20261027T140547695666Z/result.json: insert/delete costs are 6,910/6,767 instructions at 40 columns, 44,067/41,806 at 60 and 30,461/29,690 at 80. Relative to the prior checkpoint this removes 101/58, 2,293/2,258 and 2,293/2,258 instructions respectively, including the three instructions now paid to revoke the aliased index. The full 72-case matrix, five corruption classes, two cancellation cases and the registered 68-byte stack bound pass at build/styled-grid-carry-matrix/20261027T140631173980Z/result.json. Exact header and row-relative gates pass at build/lrv-header-direct/20261027T140547689491Z/result.json and build/lrv-row-relative/20261027T140547690952Z/result.json. The broad product profile passes all 24 Plain/Full fixed-grid and GEOS movement/edit cases at build/cache-hotpath-profile/20261027T141338927183Z/result.json, including every fixed-grid insert and its Undo; that lifecycle exposed the stale checkpoint alias.

Normal-speed PAL 60-column edits settle in 335,808–341,607 cycles and 80-column edits in 279,721–282,518, a 2.11–2.15% and 2.50–2.60% reduction over the parser checkpoint. NTSC settles in 338,920–344,485 and 282,018–284,822 cycles, a 1.97–2.08% and 2.45–2.55% reduction. Forty-column timing is effectively unchanged because it keeps its native text status path. Loads, calls and DMA counts are unchanged; the correctness fix increases transferred payload by 11 bytes at 40 columns and 22 bytes at 60/80. All 24 native cases preserve exact source, pixels, colors, anchors, history, catalog, payload, stack and shim, publish once, and perform no parser or overlay work after publication. Evidence and screenshots are in agentsworking/evidence/M8/20261027T140716237108Z-display-timing-styled-grid-convergence-baseline and agentsworking/evidence/M8/20261027T140850807106Z-display-timing-styled-grid-convergence-baseline. The exact D81 is c86c17087fb3d224cb6805092ce6e17695c92e91f1eaae8fd8430a532332d597.

Implemented material slice: in-place GEOS row convergence

The GEOS replacement removes repeated suffix layout and paint inside the existing C21/C22 route. It does not promote the additive R63/22-shadow prototype, add an overlay, or persist another cache. When a newly built row exactly rejoins the outgoing row stream, C21 stops the remaining C20/C24/C26 suffix construction and R28 applies one bounded signed-anchor rebase while it is resident.

The production shape uses one authenticated 842-byte following-row shadow in the existing phase loan. R28 performs one-time authority admission and publishes header VALID0. The synchronous old/new row comparator and next-row capture must remain in C21 while C20/C24/C26 and R21 run. Calling R28 again at every publish_row would reload it after those R-slot workers and fails the zero-extra-load gate before benchmarking. Replace the strict admit, verify, rebase and commit route rather than layering another transaction over it. Keep C22's existing dirty-strip admission unchanged. GEX_OLD_ROW=$D568 owns exactly $D568–$D8B1; $D8B2–$E567 stays dead and exact. C21 builds the changed row normally, then its resident comparator checks canonical NEXT, geometry, paint fields and the live CHARS/STYLES/ROLES/X planes. On exact convergence it bounds C22 to the union of old/new row strips and old/new caret strips, then terminates further C20/C24/C26 propagation. Existing publication and GCC authority remain the only commit boundary.

The final C21/R28 pair is 4,094 + 3,194 = 7,288 bytes, 118 bytes smaller than the 7,406-byte 3d53387 control. Representative Full GEOS insert falls from 907,617 to 825,542 actual 6510 instructions (9.04%), with 36 overlay loads unchanged, DMA calls falling 383 to 310 and DMA bytes falling 168,061 to 166,782. Wrapped prose insert/backspace fall about 40.4% against the original control; styled wrapped edits fall about 37.5%. The accepted transaction uses one 842-byte shadow and adds no overlay, BSS, zero-page state, REU object or persistent cache.

CPU, output and authority qualification covered all of the following:

Broad24, wrapped, complete-CR, hardening, full-page refusal, over-capacity, tail-fault, structural-refusal and Plain-mode gates pass. Lane-aware lifetime watches observed 14 accepted arms and 14 closes with zero prefix writes, owner changes, yields or open watches. C21's cached_row remains because the independent source-identical caret/selection reuse route needs it. The C32 canonical-row source check also remains as the defense against arbitrary FE CHARS corruption.

Dead GEOS edit-class path removed

The current production graph had no writer for edit class zero: canonical R28 wrote literal class five on every successful admission. C21 nevertheless kept a 142-byte class-zero plain-context synthesizer, and R28 spent five bytes writing the unread class. R28 also wrote an unread error shadow, a post-verify boundary byte and an array selector, costing another 11 bytes. All are now deleted. C21 shrinks 3,815 to 3,673 bytes and R28 shrinks 3,749 to 3,733, recovering 158 linked bytes without a new resource or relocation.

The exact ten-case wrapped control is build/current-geos-material-control-20261027T1445Z/result.json; the deletion candidate is build/current-geos-material-dead-class-20261027T1500Z/result.json. All prose, styled, Full, Markdown, INSERT, BACK and CR pixels, attributes, packets, anchors, source/history, Undo and typed-DMA masks remain exact. The complete 24-case product profile passes at build/cache-hotpath-profile/20261027T142908088409Z/result.json; the final ten-case dead-state rerun is build/current-geos-dead-state-20261027T1510Z/result.json. The deletion saves 34--38 instructions on these fallback operations because the unreachable body's entry predicates are gone; this is a code-budget cleanup, not the material convergence result. The new exact D81 is 40c59b6c09afc8db0207c74652dc0090e2b9a9f96c59f4cdfc2c571332316ddc.

Rejected local GEOS loop and repaint candidates

Two exact private candidates establish where further local tuning stops paying. The R21 compositor cleanup deletes 23 linked bytes, but saves only 17,917 instructions (0.96%) on a fresh diverse page and 2,229 (0.25%) on a styled insert. A larger role-zero shortcut is invalid: it preserves bitmap pixels but changes 26 color cells because the ninth-band color election carries histogram votes from the preceding row. Do not promote either candidate and do not spend more time on R21 leaf-loop surgery. Evidence is in build/r21-material-compositor/20261027T170000Z/current-project/build/r21-material-probe/RESULT.md.

The one-old-row C21/R28 convergence probe is exact in twelve prose/styled Full/Markdown insert, backspace and CR cases, adds no overlay loads, and shrinks the C21+R28 linked pair by 1,408 bytes against ce96920. It halves C22 calls, but saves only 319,574--344,403 instructions (18.23--21.26%) on qualifying edits and raises transferred bytes by 4,807--5,675. That misses the mandatory 500,000-instruction gate, so the patch remains rejected. Its useful result is architectural: repaint convergence is safe, but repaint alone is insufficient. The next candidate must stop C20/C24/C26 suffix construction through exact parser, carry, anchor and row-packet convergence, then reuse the unchanged suffix without adding a cache or per-row overlay reload. Report and rejected patch are in build/r28-row-convergence-probe/.

The fixed-grid packet-CRC deletion is also rejected. It removes 142 linked bytes, 22 DMA calls and about 64,400 instructions from a full publication, but the existing header key, owner, map, epoch, width and VALID flag do not prove the contents of each active FE row packet. With those authorities intact, a single corrupted packet byte is refused by the control and accepted and republished by the deletion candidate. Keep the row digest until the producer can publish an equivalent integrity witness without a second packet pass. Evidence is in build/fixed-grid-capture-tail/20261027T180000Z/RESULT.md.