Current performance
The optimization and simplification plan now governs remaining work. The63-image paint package described below is a retired experiment preserved in checkpoint6a280d8, not the accepted runtime. Later structural regression tests found a CR-deletion phase alias defect (fixed in working source) and a live narrow- glyph shadow/outgoing-checkpoint overlap (still unresolved). Its earlier scoped timings remain valid measurements of those cases; they do not qualify general editing. Committed12bde7a is the verified62-image baseline. No automatic physical Ultimate deployments are part of optimization work.
Retired experimental wrapped-page paint reuse
The registered R63 collector compares authenticated old/new row paint inputs after normal layout and semantics have run. Eligible ordinary Full/Markdown edits in wrapped GEOS source paint only the changed strip union, including old/new carets. It does not remove general layout/parsing. Plain, selection, folds, preview and unsupported geometry retain their existing paths. Every resumed collector call renews source, provider, font and publication authority; row collection is bounded to four records per call.
Fresh paired native PAL results on the ac7 raster cohort, before paged Help was combined, were:
| Full source edit | Control cycles | Paint reuse cycles | Change |
|---|---|---|---|
| Prose INSERT | 8,911,706 | 6,008,661 | -32.6% |
| Prose BACK | 8,354,393 | 5,447,900 | -34.8% |
| Styled INSERT | 10,390,956 | 8,004,637 | -23.0% |
| Styled BACK | 10,365,537 | 7,977,681 | -23.0% |
The earlier old-raster pair and this fresh current-raster pair both show the same23–35% range. All12 paired RGB images agree. CR refuses old-page reuse but still prepares the next checked publication: newlines are3.2–4.2% slower. Ordinary edits add three overlay loads; newlines add four. These costs and several-second general latency remain open, not hidden by the favorable ordinary cases. Measurements use real File/Open/IRQ setup with the disclosed public-service PC/SP stopwatch seam, not physical chords.
The combined paged-Help+paint package separately passes CPU12 edit cases, 17 authority/lifetime groups,14 per-key command transitions,29 Help/settings frames and the68-byte stack bound. Its actual2MiB native smoke checks all63 images, fontpack and five source/preview/Help frames. It is not a new paired performance run of every combined view.
Reproduce CPU product gates with tests/test_geos_paint_registered.py --fixture FROZEN63 --out NEWOUTPUT; paired native preparation/execution is in
tests/test_geos_paint_convergence_vice.py. Retained layout and complete-row
raster prototypes are independent experiments; their CPU percentages are not
current production speed claims.
Repeated eligible semantic reuse now takes6.48–6.94million PAL cycles for four successive ordinary insertions and hot-window flushes after an8KB retained code BODY. The committed010fb03 control takes9.84–10.04million:30.9–34.1% fewer cycles, with all four paired screenshots exact. C41/R34 preserve the retained BODY and reparse only the changed ordinary line (23–26 actual SuperMark feeds). The source/observer guard is renewed at each yield; up to four typed R34 operations share a synchronous call. This removes98 loads compared with the initial unbatched candidate; the final80/81 loads still exceed the control36/37.
| Native PAL repeated edit+flush+repaint | Control | Registered rebase |
|---|---|---|
| 1 | 10,044,655 | 6,939,745 |
| 2 | 9,838,779 | 6,484,229 |
| 3 | 9,848,966 | 6,497,863 |
| 4 | 9,863,435 | 6,512,032 |
These are real ROM File/Open/IRQ setup and the disclosed public-input/command PC/SP stopwatch seam, not physical-key timings. At1MHz the operation still takes several seconds. The path accepts a narrow flags5 graph and an insertion after unchanged retained code; general edits and layout retention remain open. See the authority, bounded batching and qualification contract.
Eligible ordinary GEOS Plain rows now use the existing certified local typing transaction. In paired nativePAL cases, INSERT improves2,264,434→512,064cycles (77.4%) and BACK2,254,299→180,750 (92.0%); overlay loads21/22→3/4. All four paired screenshots are exact. Same-input Full controls remain essentially unchanged. The46-frame/8-fault CPU matrix, all60-image stack analysis, eight native captures and precise REU/source/history/font guards pass. C6 and the resident are unchanged; R30/C32 grow6/34bytes. This optimizes eligible ordinary rows, with general GEOS reflow still open. See the admission, timing scope and reproduction contract.
Plain60/80 Return splits and BACK merges now use one R29 transaction when the complete affected logical paragraph is visible. Eight paired nativePAL cases improve38.1–39.9% at60 and20.4–22.4% at80, with identical screenshots and overlay loads9→4. These cases take1.02–1.45million PAL cycles; broader responsiveness work remains open. Styled, selected, folded and unknown offscreen contexts retain the ordinary renderer. See the newline transaction.
| PAL operation | Control cycles | Current cycles |
|---|---|---|
| 60 short split / merge | 1,917,609 /1,862,580 | 1,187,255 /1,145,053 |
| 60 wrapped split / merge | 2,420,160 /2,391,957 | 1,453,920 /1,438,077 |
| 80 short split / merge | 1,375,897 /1,316,037 | 1,068,295 /1,022,135 |
| 80 wrapped split / merge | 1,506,726 /1,476,201 | 1,192,321 /1,174,858 |
These compare the committed944a3e7 control with the actual cold60-image
registered product, using ROM File/Open, publicsyntax plain and the disclosed
PC/SP/public-input-to-quiescent-scheduler stopwatch seam. They include IRQ/VIC/
REU activity and do not measure physical keyboard latency. Every native VIC
pixel, source, sealed catalog asset, settings, reserved1KiB and read-only D81
passes the independent checks. CPU32paired cases separately improve29.4–38.1%
at60/18.7–26.0%at80; instruction counts are distinct from these PAL cycles.
Plain60/80 physical-row insertion/removal separately improves43.9–57.5% in paired nativePAL cases, with identical screenshots and nine overlay loads. The tested wrapped physical-row moves take about0.8–1.0seconds. See the geometry measurements and contract.
Same-call lexical batching separately improves retained-code warm PAL repaint by6.25% BASIC/6.06% assembler. It preserves286 loads and increases per-call work; long-page responsiveness remains open. See the exact paired scope.
The latest qualified80-column raster skips ink counting and color arbitration when every effective role is normal. NativePAL short-row insertion improves 255,925→226,807cycles (11.38%) and BACK255,319→226,366 (11.34%), with four overlay loads unchanged and identical screenshots. Wrapped-quote typing remains essentially unchanged at365,145/368,335cycles. This is an ordinary-row improvement; whole-editor responsiveness and large GEOS work remain open. See the implementation and exact qualification.
These measurements compare the946e7d3 control with the same59-asset product plus
R16 and its genuine R25/C33 CRC witnesses. They use public syntax plain, ROM
File/Open and the documented PC/SP/input-entry stopwatch seam, not physical-key
latency.934component rows plus17invalid inputs per variant,40 public comparison
frames,20 wrapped frames,24 refusals, canonical editor/cache and complete68-byte
relative stack checks retain their exact scopes. Normal componentrows use30.77%
fewer instructions; mixed-role fallback costs1.36% more at that component level.
Earlier54-asset checkpoint
The earlier cache-redesign runtime is 69720da, with 54 assets, resident
a2578487c8a5 and D81 2ffdb2fc95b1. Its diverse-document full GEOS repaint
was measured at 10,150,476 PAL cycles in Full and 9,629,277 in Markdown. All
eleven native display captures match the reviewed images. Full still exceeds
the older 9,511,774-cycle path by 6.7%; performance qualification remains open.
These timings were captured on 4c25480; the subsequent C19 pressure fix
changes only fold allocation and is qualified separately below.
A separate native control loads a 4,102-byte document through File/Open: a long bold span followed by a short paragraph. These public commands include actual navigation and repaint, after the semantic baseline has been built:
| PAL command | Control cycles | Current cycles | Loads, control → current |
|---|---|---|---|
| Markdown bottom | 306,428,091 | 4,087,703 | 577 → 32 |
| Plain bottom | 14,247,606 | 14,229,896 | 176 → 176 |
| Markdown top | 28,352,606 | 22,550,943 | 243 → 249 |
The actual dispatcher trace changes from 13,598 C26 requests and 223 C24 requests during styled bottom to two C26 requests, fifteen bounded C35 steps, and no C24. The unfolded jump queries its target's retained block state and starts measuring at that block's logical start. It skips preceding blocks; softwrapped rows inside the target block still require measurement. Ordinary UP/DOWN/PAGE/PAN and cached geometry retain their existing paths. All three current images match the control byte for byte. Bottom improves 98.7%; Top improves 20.5% despite more loads. Cache hits and parser-feed counts cannot replace end-to-end timing. The measurements use the documented public-input PC/SP seam and include IRQ/VIC/REU activity; they do not measure physical keyboard latency. All three images, source/history/clipboard/settings, retained typed-cache bytes and the full REU outside checked renderer storage pass verification.
R21 now intersects the glyph's vertical clip interval once per strip rather than testing it on every scanline. The same packed-cold control improves from 10,286,041 to 10,150,476 PAL cycles (1.3%) with identical pixels and 149 loads. The payload remains 4,094 bytes; fixed font data stays identical. Component coverage includes 2,570 vectors, 65 faults and 720 legal vertical intersections per side. This improves raster work without claiming to solve overlay traffic.
The latest C19 fix adds 99 bytes and no resident state. Fold allocation retries once after retiring the optional checked type53 cache, only for actual object catalog or provider exhaustion. All 28 fold scenarios, 14 allocation and ownership cases, five editor scenarios and the 68-byte stack bound pass. A new native Copy/Cut/Undo gate checks three frames, 192,000 VIC pixels, exact source/clipboard/folds, all 54 assets and the 1 KiB reservation. TRIED remains set after retirement, so automatic cache rebuilding stays deferred; normal parsing continues.
Current evidence: agentsworking/cache-raster-vertical-checkpoint.json and
agentsworking/cache-fold-pressure-integration-checkpoint.json. The native
pressure captures completed successfully; a host module-alias error was repaired
offline against those same captures, with the failed log retained.
Historical navigation integration evidence: agentsworking/cache-navigation-integration-checkpoint.json and
agentsworking/evidence/M8/20261026T045738645714Z-navigation-cache-integration.
The control archive retains an initial host-evaluator capture-name error and
its repaired offline verification. The new production run passes directly.
Packaged navigation56, parser bypass55, incremental15, stack68, editor5 and
foreground2 pass. Retained lexical colors and edit-aware semantic rebasing
remain private experiments; these measurements do not include them.
Earlier 48-asset checkpoint
D81 dbc4ecf968ab, unchanged resident2d87e75e/48assets. geoWrite-style
complete logical-line convergence and source pixel batching are integrated.
The goal remains active; these are incomplete performance improvements.
| Current PAL operation | Cycles |
|---|---|
| Diverse GEOS heading-inline INSERT /BACK | 2,557,495 /2,011,074 |
| Fresh GEOS diverse TOP0 repaint | 9,500,660 |
| Ordinary heading insertion | 2,206,795 |
| Source60 /80 mixed fallback typing | 10,557,053 /7,845,115 |
| Warm RIGHT40 /60 /80 | 12,916 /132,700 /137,077 |
| GEOS warm LEFT, zero loads | 36,670 |
| Preview first /unchanged repaint | 7,894,608 /142,368 |
The isolated line-convergence A/B changes the diverse INSERT from9,814,874 to2,551,898cycles with identical pixels; the combined current value above includes separately improved C20. Source paragraph native12 on0340 shares exact current C14/R14/C33/R25: long60 .920–.935M, long80 .841–.859M cycles.
Regressions are retained: fullTOP0 is slower than6604's9,077,527PAL; ordinary heading typing is slower than1,984,869. Isolated carry/wrap fallback adds 247,318/116,357cycles. PAL32 pixels exactly match6604; fresh NTSC32 passes with matching320x200 crops. An earlier preview-page failure during host ENOSPC is retained; cause not assumed. Fresh fullTOP0 NTSC is9,579,924cycles. Do not treat these changes as general typing/usability closure or a geoWrite speed benchmark. Reuse still requires a complete single-visual-row logical line; wrapped convergence is private work. Screenshots, exact-build scope and receipts are in STATUS.
Historical6604 measurements
Earlier D81 66045b585ddb, resident 2d87e75e,48 assets/three asset banks.
The standalone goal remains active: full redraws and several editing paths
still miss the responsiveness targets.
| Actual operation | PAL cycles |
|---|---|
| Fresh GEOS diverse page, TOP0, warp off | 9,077,527 |
| Same TOP0 at preceding e34 checkpoint | 10,365,612 |
| Warm RIGHT,40 /60 /80 columns | 12,662 /132,630 /136,879 |
| GEOS warm LEFT, no overlay loads | 36,671 |
| GEOS first RIGHT, one optional AUX load | 41,274 |
| Diverse GEOS heading insertion | 1,984,869 |
| First formatted preview | 7,884,262 |
| Unchanged preview repaint | 141,892 |
| Mixed source60/80 insertion, settled | 10,831,796 /8,119,655 |
Fresh TOP0 is about12.4% faster than e34, but remains roughly nine seconds at normal C64 speed. NTSC is9,152,893cycles. Matrix full paint uses TOP318 and costs3,731,541PAL: it must not be compared as the same page. The measured private fixed-shift change alone contributed only0.26%; avoiding repeated glyph/frame work and style parsing accounts for the larger combined improvement.
The source60/80 mixed fallback above regresses from roughly9.46M/6.75M before SP2. Its new complete-paragraph authority checks cost work even on rejected paragraphs; reducing that cost is active work. Eligible long wrapped prose now uses a bounded pixel/packet transaction, around0.9–1.0Mcycles in the exact a6fa native8 proof. Its C14/R14/R25/C33 images are byte-identical in current6604; only the GEOS C20 worker changed afterward. That is still not ordinary typing performance. Markup, selections, folds and geometry changes use checked fallbacks.
C20 reuses its completed output in C21, and retains a160-byte style window during a line job. Current-binding CPU A/B versus the first certificate saves1.44% on mixed422,6.03% on inline406 and26.94% on long styled863. Styled wideW200 costs 3.52% more; short cases add less than0.07%. Source reads, full-body delimiter semantics and source identity checks are preserved. There is no cross-line-job style reuse yet. CPU instructions are not native cycles or DMA byte totals.
Actual after-publication parser-checkpoint refresh prevents repeated lower-row edits from aging otherwise useful incoming state. The73-frame public test plus 256 readonly publications verifies refresh, rare lowbyte collision, missing workers and actual Undo recovery. Full16-bit/source/CRC checks remain inside AUX; the lowbyte resident gate is never treated as authority.
Current public15/inline19/fold28/capacity12/recovery22 and PAL32+NTSC32 pass. All32PAL PNGs match reviewed ancestors; NTSC active crops equalPAL. Full48-image stack remains77/minSPB2 and forced66-output rebuild is byte-exact. Focused normal-speed intervals bracket warp-off monitor witnesses; boot/setup retain warp and drive traps. Public-entry PC/SP stopwatch seams include real scheduler, IRQ,VIC and REU costs, but do not measure physical keyboard latency.
Evidence: agentsworking/evidence/M8/20261025T205616461479Z-combined-style-refresh-production/manifest.json.
Final combined lifecycle, media, REU-pressure and gallery acceptance remain
pending alongside responsiveness work. Historical proofs retain their scope.
Previous renderer checkpoint
Current48 renderer build: resident b33cc41e, D81 57be0ca4350b,48 assets.
Strip clipping, bounded packet gathering and removal of unused row snapshots
are promoted. No extra REU bank, resident state or zero page is used.
| Same diverse document / operation | PAL cycles | Loads |
|---|---|---|
| 40-column RIGHT | 12,900 | 0 |
| 60-column RIGHT | 134,657 | 2 |
| 80-column RIGHT | 138,740 | 2 |
| GEOS heading-prefix insertion | 2,315,926 | 17 |
| GEOS fresh full page, TOP0 / warp on | 11,262,377 | 118 |
| GEOS fresh full page, TOP0 / warp off | 11,262,403 | 118 |
| GEOS fresh full page, TOP318 | 4,614,962 | 50 |
| First formatted preview | 7,771,315 | 67 |
| Unchanged preview repaint | 139,978 | 5 |
TOP0 full paint improves about8.1% from the same12.258M-cycle viewport in the preceding checkpoint. TOP318 contains different content and is not its comparator. Focused NTSC TOP0 is11,355,937 cycles with warp on versus11,355,981 with warp off. Each normal-speed interval has two actual monitor witnesses that warp is off; setup/boot use warp and drive traps. Timing runs preserve the disclosed public-entry PC/SP seam and actual IRQ/scheduler/VIC costs. They do not measure physical key latency or accelerated hardware.
All64 packaged PAL/NTSC native images match reviewed ancestors/active crops. Actual CPU heading72, boundary136 and recovery/history gates preserve source, fonts, pixel/color output and checked REU ownership. Current full stack remains 77bytes/minimumSP B2; forced rebuild reproduces65 outputs byte for byte.
General60/80 typing, GEOS fullpage/CR/softwrap reflow and far-prefix navigation
remain slow. The private short-prose typing and further raster-loop candidates
are not production claims. Lifecycle/media/capacity evidence is inherited in
its declared unchanged-contract scope. Evidence:
agentsworking/evidence/M8/20261025T182719631618Z-renderer-production/manifest.json.
Historical combined48 before renderer deltas
Current48: resident b33cc41e, D81 da5f2235a110,48 assets. The integrated
build combines retained60/80 caret deltas, heading-body row rebuilds and exact
shared-tile color election. Measurements are emulated C64 cycles including
IRQ/VIC activity, with warp/traps enabled and a disclosed public-input PC/SP
stopwatch seam. They are not physical keyboard latency or hardware acceleration.
| Integrated diverse document, PAL | Cycles to settled idle | Loads |
|---|---|---|
| 40-column RIGHT | 12,899 | 0 |
| 60-column RIGHT | 134,442 | 2 |
| 80-column RIGHT | 138,745 | 2 |
| GEOS RIGHT after mode entry | 41,245 | 1 |
| GEOS sustained RIGHT | 36,504 | 0 |
| GEOS full top page, TOP0 | 12,257,971 | 118 |
| GEOS heading-prefix insertion | 2,506,384 | 17 |
| First formatted preview | 7,771,418 | 67 |
| Unchanged preview repaint | 140,194 | 5 |
The preceding60/80 RIGHT costs were875,167/795,756 cycles on the same source. C33 now changes only certified caret/selection cells and status. Its private native29 and timing10 additionally cover cross-CR movement, selection, menus, File owner/STOP and folded planning; source40 retains zero-load navigation.
Private exact-algorithm qualification measured heading-body insertion in the 422-byte diverse document at12,450,919→2,428,706 cycles, and BACK at 12,318,033→2,296,549. Natural prose below a heading measured247,704 cycles with no overlay load; a longer pangram340,239. These private measurements have separate identities; the combined48 CPU gates verify integration. Unsupported color-winner changes and the minority-color counterexample still use full rendering. General CR/softwrap reflow and long-prefix parsing remain slow. The heading candidate also adds about458k cycles to one measured new-wrap refusal; correctness holds, but that regression remains optimization work.
Current packaged PAL32+NTSC32 pass64 independent native images/4.096M pixels;
all PAL images equal reviewed predecessors and NTSC active crops equal PAL.
The original matrix's source-GEOS m3-full retained a caret-cache hit: its
38,893-cycle cost is not full paint. The canonical benchmark now invalidates
the authenticated GEOS bitmap/edit witnesses and width before source full paint;
focused corrected PAL/NTSC measurements are12,257,971/12,360,021 cycles at
TOP0. The earlier4.915M full figure used TOP318 and is a different viewport,
not a before/after comparator. Preview hit remains a
separate operation. Forced rebuild reproduces65 existing outputs exactly.
Evidence: M8/20261025T171958460619Z-combined48-production.
Historical47 span-color checkpoint
Historical47: resident 4e64d113, D81 ed098bf5c005, 47 assets. These are
emulated C64 cycles with IRQ/VIC activity, warp and drive traps. Timing uses a
disclosed public-input PC/SP seam through real foreground quiescence; it does
not measure physical keyboard latency or accelerated hardware.
C14 source coloring now seeks visible row intersections instead of scanning all screen cells for every span. The promoted payload is4047 bytes (+124), with no new state. Exact actual-root and native A/B checks preserve all colors, source/history/REU and ownership behavior. Settled PAL costs on the same edit:
| Source mode | Before | After |
|---|---|---|
| 40 columns | 11.551M | 2.947M |
| 60 columns | 22.340M | 9.463M |
| 80 columns | 23.726M | 6.750M |
The first publication is essentially unchanged; these figures measure less background styling. The current packaged PAL+NTSC32+32 also passes64 frames, with unchanged reviewed raster output. All64 existing build outputs reproduce exactly. Evidence: M8/20261025T162303429392Z-style-span-production and M8/20261025T161651340930Z-style-span-editor-candidate.
The local GEOS styled-edit numbers below retain their measured893dc fixture; its R30/C32 and all other46 assets are byte-identical to currented098. C14 does not participate in those local GEOS transactions.
| Current shared styled/plain document | PAL cycles | Overlay loads |
|---|---|---|
| Cold first bold insert | 608,368 | 1 |
| Italic insert after font-window refill | 491,298 | 1 |
| Combined insert with retained staged code | 193,007 | 1 |
| Bold / italic / combined BACK | 170,137 /152,119 /157,164 | 0 |
| Plain BACK | 152,849 | 0 |
| Wide bold W insert / BACK | 179,811 /152,552 | 1 /0 |
| Wide combined M insert / BACK | 192,814 /157,862 | 1 /0 |
Both PAL and NTSC pass all12 captures and six complete16MiB/source/history intervals; all current PAL PNGs match their individually reviewed same-document predecessors. NTSC active pixels match PAL. C32 splits advances8–11 into one byte plus0–3 bit shifts. This adds149 bytes/no state. Warm wide deletion saves 19.2k–26.7k PALcycles; cold prime adds about13.7k, while narrow edits are nearly flat. These costs include staging and validation rather than just the shift loop. Evidence: M8/20261025T140416711643Z-style-typing-profile-native and M8/20261025T140623218510Z-style-typing-ntsc-native.
The current full32 PAL+NTSC matrices also pass64 independent captures. The corrected fresh-page seam measures source-GEOS full paint at4,915,488 PALcycles; clearing only the old grid-valid bytes previously allowed a38,741-cycle caret cache hit, which must not be called full paint.
| Current diverse workload | PAL first publication | PAL settled idle |
|---|---|---|
| 40-column insertion before heading | 279,733 | 11,551,172 |
| 60-column warm RIGHT | 875,368 | 875,476 |
| 80-column warm RIGHT | 795,995 | 796,103 |
| GEOS fresh full paint | 4,915,488 | 4,915,572 |
| GEOS mixed prose insertion below heading | 2,666,197 | 2,666,281 |
| GEOS heading interior insertion | 12,328,587 | 12,328,671 |
| GEOS CR line split | 12,159,611 | 12,159,738 |
| GEOS real softwrap314+8→322pixels | 7,764,343 | 7,764,427 |
First publication and settled styling are separate: input may be processed while background styling progresses, so these are not interchangeable with physical key latency. The60/80 caret rescan and slow GEOS structural paths are remaining implementation problems. The mixed-prose target is itself plain; a colored heading shares its physical8px tile band. Its fallback preserves minority-color correctness; simply removing that guard would reintroduce a proved wrong-color regression. A bounded color-vote repair is future work.
All-mode measured intervals total651 loads,29,609 overlay calls,3444 DMAs and 2,660,651 DMA bytes per standard. These are sums of32 measured windows, excluding setup/boot. Full tables, exact IRQ/trap/seam/guard scopes and reviewed image ancestry: M8/20261025T142349107759Z-current-all-mode-timing. Diverse7: M8/20261025T141546628739Z-display-timing-pal. Real softwrap: M8/20261025T141826196287Z-display-timing-pal. Native source/asset/pixel guards are retained; those gates do not assert arbitrary whole16MiB immutability.
The immediately preceding D81 46b137568169 has the same resident and45
identical assets; only gtype/type differ. Its unchanged-renderer measurements
remain scoped to that disk:
| Event on diverse document | PAL cycles | Loads |
|---|---|---|
| Warm caret left | 36,501 | 0 |
| Unchanged preview repaint | 140,035 | 5 |
| First preview | 7,769,787 | 67 |
| Return to previous preview page | 6,758,066 | 67 |
| Single alpha before a complete heading | 2,411,793 | 17 |
Warm caret updates old/new tiles and one status row. Certified typing shifts only its row suffix; code edits repaint a complete short row with the actual lexer. Neither requires a full page. Narrow heading transitions rebuild the actual changed row before reusing the tail; general structural edits still use checked full reflow. First preview and general reflow remain slow. Current all-mode scrolling/full-redraw and diverse reflow costs are measured above; end-to-end file open/save costs are recorded below. No universal latency bound or completion of the provisional frame/redraw targets is claimed.
Managed preview-body reuse cut an unchanged repaint from1,488,860 to378k cycles; chrome batching then cut378,176 to140,035. Body/source/settings/font/fold/FE guards remain, and checked writes revoke before hardware publication. R23 caches compact BSW9 glyphs in4160 bytes of the existing typedFE tail; immediate wrapper fetch reuse retains the guarded direct API. First preview fell8.51M→8.01M before chrome batching. Early FE allocation was deferred because empty/short pages regressed despite some larger-document gains.
Chrome rows0/1/24 use three checked row calls instead of120 cell calls. Popup
rectangles retain the cell API. C16=4096/R27=4053; no new BSS/ZP. STATUS adds167
modeled instructions, a small caret regression (36,021→36,501 in the measured
PAL fixtures). Source/history/16MiB/stack/shim guards remain. Preview evidence:
M8/20261025T135131738122Z-display-timing-pal and
M8/20261025T135321963581Z-display-timing-ntsc.
The full scoped checkpoint, including reproducible64-output build and native
gallery43/F8five/lifecycle6 ancestry, is linked by
agentsworking/byte-render-checkpoint.json.
The following earlier measurements retain their historical identity.
Latest47 local-style increment: resident d69f4f0f, D81 a49e46c559e0.
The following newer measurements use one shared styled/plain document and
supersede only the corresponding local-typing results, not other workloads.
| Actual event | PAL cycles | Overlay loads |
|---|---|---|
| First bold insert, cold R30 and staged capsule | 601,770 | 1 |
| Italic insert after a font-window refill | 484,149 | 1 |
| Combined-style insert with retained staged code | 212,508 | 1 |
| Bold BACK | 189,446 | 0 |
| Italic BACK | 193,518 | 0 |
| Combined-style BACK | 204,063 | 0 |
| Plain BACK in this shared fixture | 172,687 | 0 |
| Wide bold W insert/BACK | 199,302 /191,870 | 1 /0 |
| Wide combined M insert/BACK | 212,329 /204,051 | 1 /0 |
These12 native frames pass exact source/history and six full16MiB intervals.
All images are byte-identical to the individually reviewed preceding47 frames;
the new bold/combined images were also directly viewed. Evidence:
agentsworking/evidence/M8/20261025T115900566566Z-style-typing-profile-native.
C6 currently requires an unchanged alphabetic line opener and certifies
conservative alphanumeric interiors; a line beginning with a style marker can
still use the slower fallback. R30 authenticates and runs a
C32 transaction from5000. It uses real bold/italic ink and updates full packet
styles/X geometry. Each admitted edit patches its row, with no page, semantic,
inline, source-provider or raster producer calls. Spaces and markup edges can
still fall back; this is not a universal latency bound.
The22-byte suffix-bounds optimization cuts warm styled BACK from238–279k to 189–204k cycles on the same fixture. It preserves unused-tail clearing and shared-tile color checks. Further transfer and preview work remains private.
The following managed46 measurements remain exact historical comparisons.
Measured in native VICE with IRQ/VIC activity, warp and drive traps. These are
emulated PAL C64 cycles to completed bitmap, attributes and status, not physical
keyboard latency or accelerated-hardware measurements. Current build:
resident bdef9dd0, D81 b3813f5d4fae,46 images.
| Operation and workload | PAL cycles | Overlay loads |
|---|---|---|
| Warm GEOS caret LEFT, diverse422-byte document | 35,984 | 0 |
| First caret RIGHT in that profile | 40,317 | 1 |
| First ordinary insert,61-byte/four-row document | 454,941 | 1 |
| Warm ordinary insert, same document | 189,316 | 0 |
| Warm ordinary BACK, same document | 186,450–206,496 | 0 |
| Insert/BACK in its second row | 205,072–212,972 | 0–1 |
| First BASIC string insert, direct R31 | 641,460 | 1 |
| BASIC string BACK, direct R31 | 128,829 | 0 |
| Assembler comment insert/BACK, direct R31 | 407,265 /111,692 | 1 /0 |
| BASIC number insert/BACK after retained stage | 116,768 /111,714 | 1 /0 |
| Structural/styled insert requiring full redraw, diverse document | 13,711,876 | 120 |
| First formatted preview, diverse422-byte document | 8,756,113 | 67 |
| Next preview page | 7,737,297 | 77 |
| Return to first preview page | 8,299,256 | 67 |
| Unchanged preview repaint using body cache | 1,488,845 | 5 |
Warm caret updates only old/new caret tiles and the single status line. Ordinary admitted typing shifts the affected bitmap suffix, updates its row packet and later source positions, then updates caret/status. It does not run the source, semantic, inline, page or raster producers. Current plain timing includes the neighbor-role guard: every glyph intersecting the physical tile bands must be plain before colors can remain unchanged. A white displayed attribute alone is insufficient. Other cases retain the checked color-voting fallback.
R31 recomputes a complete short aligned BASIC/assembler row using the actual
lexer and native8 painter in sealed C31 code at $5000. Warm edits retain it.
The six-case profile deliberately includes different first-use states: its
BASIC-number insertion benefits from a previously verified overlay and retained
staged code. It is not a universal cold-load result. All six actual publications
match independent full pixels, source/history and complete REU guards.
The editor now remembers verified immutable overlay assets and copies warm images directly to their execution slots. This removes repeated payload CRCs and preserves the font window. Imports/unverified loads still check fullCRC; every entry checks its resident binding. Checked code writes, provider changes and fresh resume revoke verification. The generic loader remains strict.
On the same preview fixture, first preview improves from21,689,902 to8,756,113 cycles (60% reduction); unchanged repaint improves from2,462,433 to1,488,845 (40%). The screenshots are byte-identical to the previously reviewed images. This is useful progress, but a roughly8.9-second cold preview at standard PAL speed and the structural redraw above remain performance blockers. Warm plain editing is roughly0.19–0.21seconds; it also exceeds the original provisional six-frame typing target. These scoped improvements do not close M8.
The body cache stores7040 bitmap bytes and880 attributes in the existing typed FE allocation. It verifies source/settings/folds/fonts/provider context and full bodyCRC before two live display DMAs; no semantic/inline/lexer/font worker runs on a hit. The retained font fast path relies on the actual checked window capability, not the number of unrelated overlay loads.
Evidence:
agentsworking/evidence/M8/20261025T112838129760Z-display-timing-pal: seven preview/caret/structural publications and cost profiles.agentsworking/evidence/M8/20261025T112849904736Z-incremental-profile-native: six ordinary edits, paired full16MiB REU intervals and source/history checks.agentsworking/evidence/M8/20261025T113512024238Z-code-typing-profile-native: six BASIC/assembler edits, three paired whole-REU intervals.
Every listed screenshot was visually inspected. Native timing enters the public foreground service through a disclosed measurement seam; separate lifecycle and editing automation proves actual IRQ-delivered typing. CPU pixel gates compare the optimized editor with a forced actual full renderer and an independent source/font oracle. Physical keyboard and Ultimate acceleration remain separate.
End-to-end file costs on current893dc
PAL modal Open422B settles in12,444,142 cycles; SaveAs423B in14,478,856; a separate fresh-process reopen423B in12,446,796. Their first publications are about1.20M/3.21M/1.20M respectively. Settled costs include rendering and background styling; these are not isolated IEC transfer rates. Exact saved SEQ bytes and fresh reopened source agree, alongside independent native pixels, owned catalog/payload/fontpack/history/savepoint/shim/lowerstack/probe guards. The read-only reopen disk is unchanged. Warp and device8 traps are enabled; PAL VIC0/50Hz resources and a16-cycle calibration are retained. Evidence: M8/20261025T142629350991Z-file-throughput-native. No physical latency or worst-case claim is made.
Warm GEOS byte shifts on the registered62 editor
A bounded bitmap suffix change removes the second neighbor from1..7-bit shifts and performs an8-bit insertion shift as a byte displacement before its0..3-bit remainder. The payload is3 bytes smaller (C32 raw3771, sealed3803); R30 remains 3662 bytes with updated stage length/CRC constants. Admission, color voting, source/history certificates and publication checks remain in place.
Actual PAL native measurements through the disclosed public service PC/SP stopwatch, with IRQs running:
| Case | Control cycles | Optimized cycles |
|---|---|---|
| Plain steady INSERT | 175856–175890 | 169637–169888 |
| Plain BACK | 170707–170928 | 164599–164621 |
| Full strong steady INSERT | 186356–189569 | 178295–181178 |
| Full strong BACK | 181049–184271 | 173407–176108 |
| Plain first INSERT | 717300 | 711039 |
| Full strong first INSERT | 515393 | 507786 |
Steady keys improve3.4–4.4%; first insertion improves0.9–1.5%. Caret code is unchanged and its IRQ-phase variation includes a0.6% regression. The48 native captures have24 byte-identical control/candidate PNG pairs and8 distinct images, all visually reviewed. Both packages retain actual cold62 registration, fonts, source/history, settings, clipboard, the reserved shim, and precise whole16MiB REU checks. The provider is explicitly configured as256 banks. These timings measure service cost rather than physical keyboard latency.
The native pair keeps the same earlier C41 in both variants. The combined 7654 correctness fix plus new TYPE/GTYPE pair is separately authenticated and passes canonical rebase protocols/faults, real24-COW journal boundaries, public pixel/authority gates, complete62-image stack analysis and no-op build checks. Only those two packaged images differ from7654. Evidence: native build/warm-shift-native-adapter/ready8 and canonical build/warm-shift-canonical-{protocol,guards,journal,stack}; source-matched working fixture build/warm-shift-canonical-working-v2/20261027T001110182649Z/fixture. The provisional overall latency targets and broad redraw gate remain open.
VIC scatter in groups of eight columns
A second warm-typing change replaces40 pointer updates per raster line with five8-column groups. It retains all40 writes, nine raster lines, original tail zeroing, caret-free base, color rules, source certificates and publication order. The staged payload grows53 bytes to3824 (sealed3856); R30 remains3662 with its regenerated length/CRC witness. No state, allocation or overlay is added.
Actual native PAL service measurements, committed1612523 control versus the coherent two-asset candidate, both with the corrected C41:
| Case | Control cycles | Optimized cycles |
|---|---|---|
| Plain steady INSERT | 169627–169856 | 163818–163859 |
| Plain BACK | 164597–164622 | 158693–158927 |
| Full strong steady INSERT | 178297–181173 | 172983–175726 |
| Full strong BACK | 173393–176320 | 167661–170640 |
| Plain first INSERT | 710767 | 709393 |
| Full strong first INSERT | 507786 | 506770 |
Steady keys improve3.0–3.6%; first insertion improves about0.2%. Caret code
is unchanged; three measured caret intervals regress by1/6/4 cycles, retained
as IRQ-phase variation. All48 actual captures agree with independent native
pixels and24 paired activeRGB comparisons. Their eight distinct PNGs hash
exactly to previously root-reviewed originals. Configured provider256 and tight
whole16MiB guards, font/catalog/source/history/shim and actual stage capability
checks remain. Source-rebuilt coherent D81 matches the canonical package byte
for byte, not merely its logical files. Native evidence:
build/vic-stride-probe/native-ready3; CPU differential/negative evidence:
build/vic-stride-probe/receipt.json. General redraw and latency targets remain
open. CPU instruction savings4.48–5.18% are deliberately separate from native
cycle savings3.0–3.6%.
Queued input cancels an obsolete GEOS repaint
The foreground scheduler now cancels source GEOS page job11 when an input event is already queued, using the existing lease/cancel path and leaving that event for normal input dispatch. Job12 still completes requested navigation before an ordinary queued edit. This replaces waiting for a required page paint to finish; there is no new overlay, state allocation or parser cache. Resident code shrinks one byte after moving the shared dispatch branches.
A paired PAL VICE check starts with a genuine pendingC22 sourcePlain repaint of1890 bytes, enters Q through KERNAL/IRQ, and times from the stopped foreground boundary with that event queued until it is applied. Actual pending/queued/ applied RAM states are asserted; neither PC/SP nor job/cache/source is seeded. The verified62 control takes9,290,657 cycles; the candidate takes8,237. Both produce identical final bitmap/color cells, independently checked source and 64,000 native VIC pixels, with the1KiB ReadyOS reservation intact. This is queued-event service latency at a safe boundary, not physical key-to-photon or a bound on an already executing render slice. Warp/traps accelerate the host run; the measurement uses emulated PAL CPU cycles with IRQ/VIC/REU stalls.
Ten candidate CPU checks and six paired repaint checks cover Plain/Markdown/Full pendingC20/C22, a16-event burst, requested-navigation ordering,STOP/NMI cleanup and subsequent40-column recovery. The six repaint cases reduce queued apply137,267–3,297,831 to1,279–2,131 CPU instructions with exact final pixels and Undo/Redo. These are instruction counts, not cycle estimates. Actual62 foreground+IRQ/NMI stack bound remains68 bytes. Targeted preview checks confirm that a queued second page command waits for the completed bottom anchor, with original STOP cleanup/recovery retained. Formatted job12 now also retains its destination while its semantic worker is active: the old global cache-cancel branch could insert at the old caret. The larger renderer simplification remains open; this does not close general 1MHz responsiveness.
Evidence: build/queued-repaint/final-v3.json, build/queued-repaint-supplement/v3-small-final/result.json, tests/test_geos_queued_repaint.py and tests/test_geos_queued_repaint_vice.py. Initial native v2 timing receipts are explicitly rejected: an asynchronous host playback script lacked its completion sentinel. The v3 comparisons assert every timing boundary from captured RAM.
Compose a logical GEOS row once
R21/C22 gather and compose a complete logical row instead of restarting its preparation for each8-pixel physical strip. Existing fonts, color arbitration, selection, clipping and native code rows remain. The current pair is4095+4089 bytes versus4094+4056:34 extra payload bytes, no new image or resident allocation. The private whole-row packet uses existing exclusive phase loans.
The renderer polls the resident IRQ ring/NMI after initialization, at most four complete glyphs, and each40-tile resolver band. If no event is pending it continues within the same verified call. Pending input returns immediately without consuming it; the scheduler cancels obsolete source paint before applying the event. Source/provider/header/selection guards remain at actual public STEP boundaries. This avoids repeated scheduler/validation work just to resume a row with no input waiting.
Paired native PAL File/Open→Bottom→F7 Top commands use the established final Return public-input stopwatch seam with PC/SP setup. They measure complete command work, not physical key latency. Independent source/bitmap/colors and actual native VIC pixels agree.
| Public Top command | Original cycles | Current cycles | Reduction |
|---|---|---|---|
| 863-byte Full prose | 15,422,210 | 13,220,193 | 14.28% |
| Wrapped Markdown prose | 11,500,650 | 9,713,018 | 15.54% |
Loads remain118/24; total time is still far above the original1MHz goals. Without waiting input, full service invocations can now reach598,490/562,228 cycles because work continues across internal polling points. This is not a bound on how long an arriving event waits.
A separate actual IRQ check starts the stopwatch immediately after the ring publishes Q while R21 is drawing. The saved ROM IRQ stack proves the interrupted PC is inside the active R21 image; no PC/SP/job/source is seeded. Publication to applied input falls51,447→21,026 PAL cycles versus the initial eight-glyph version. Final pixels/source/shim agree. This one arrival point is below50k; universal latency across every clip/font/IRQ phase remains unproven.
Private CPU12 cases all improve versus the old strip renderer: long warmed Full−20.65%, diverse Full−5.96%, diverse Plain−0.88%. Clipping48, faults49, asynchronous arrivals12, incremental15 and warmfold28 preserve independent pixels/source/history/shim and precise wholeREU guards. Actual integrated queue10 and62-image foreground+IRQ/NMI stack68 pass. Historical cold-fold audit still rejects initial type53 writes on both cohorts; warmfold explicitly completes that real baseline without relaxing its evaluator.
Evidence: build/material-native/poll-qualified.json and build/row-residency-probe/READY-POLL4.json. Native recipes are tests/test_geos_material_profile_vice.py and tests/test_geos_queued_repaint_vice.py --inflight. This replaces the initial 9bf262a unconditional eight-glyph continuation; full final acceptance stays open.