ULTIMATEEXPERIMENTSCommodore 64 · Work in progress

Plain 60/80-column editing

syntax plain now uses the checked grid packets for caret movement and local typing without first loading semantic parsers. This changes three existing capsules: C33 grid, R25 geostail, and R14 viewsource. That checkpoint kept the resident, allocations, zero page and ReadyOS reservation unchanged. The subsequent lexical-cache registration brings the catalogue to59 entries.

NC header byte253 records the exact presentation preset under the existing CCITT checksum. Every consumer compares it. Plain typing checks source objects, window/revision/caret deltas, exact old-to-new bytes, packet CRCs, row ownership, fonts and REU capabilities. It never manufactures an SP2 semantic certificate, publishes a semantic descriptor or marks parsing valid. Markdown and Full retain their semantic checks and exact helper witnesses.

Short edits require both old and new row lengths below the column width. Wrapped edits can carry pixels across existing rows only after validating the complete affected source and packets. Before writing pixels, a real checked R14 admission authenticates its pixel worker. Its owner4 operation6 is a no-op used to complete that admission; operation7 performs the existing pixel carry. Four assembler- checked short branches fund the new entry without removing validation.

The short-row old-length check is essential: inserting into a width-minus-one line can create a separate CR row. Deleting the inserted byte must remove that physical row. An earlier candidate checked only the new length and left stale following pixels. The native test reproduces this sequence in both widths; the corrected build uses the full renderer for this geometry change.

Measured PAL work

These are VICE CPU cycles through settled foreground idle, with IRQ/VIC/REU stalls included. Setup uses real ROM File/Open and IRQ commands. The stopwatch uses an explicit PC/SP/public-input entry seam, so these are not measurements of physical keyboard latency. Warp and drive traps are enabled for development.

Plain operation Previous Current Current overlay loads
60-column Right 871,683 133,289 2
80-column Right 792,692 137,911 2
60-column short-space insertion 997,820 371,328 4
80-column short-space insertion 931,728 255,659 4
60-column first wrapped-quote insertion 1,435,327 692,236 4
80-column first wrapped-quote insertion 1,029,577 364,769 4

Structural edits still have a material regression because their completed full paint is followed by the 22-packet capture:

Plain fallback Previous settled Current first publication Current settled
60-column insertion adding a row 1,428,082 1,523,967 2,046,346
80-column insertion adding a row 1,125,862 1,221,684 1,740,134
60-column CR insertion 2,094,539 2,243,720 2,760,409
80-column CR insertion 1,538,007 1,691,286 2,205,244

The current capture implementation reduces the extra settled work from roughly 0.52 million to0.30 million PAL cycles. It expands full CRC lookup tables once and reuses the already admitted R25 during one exclusive capture. Every helper still checks its actual owner/operation; no executable image is patched. Both pending redraw and drawbusy refuse before any borrowed scratch is initialized.

The lookup tables alias retained line-index storage. Capture now revokes all three index/cache authority fields before writing there. The first native mode60→80/GOTO2 test exposed a $7B error when that invalidation was missing. Both mode directions, real File/Open, and GOTO above64KiB now pass.

Current fallback First publication Settled PAL cycles Loads
60-column insertion adding a row 1,524,117 1,828,709 9
80-column insertion adding a row 1,221,519 1,525,607 9
60-column CR insertion 2,243,693 2,546,706 9
80-column CR insertion 1,691,372 1,994,023 9

Handling changed geometry efficiently remains open. Full GEOS source repaint is still10,155,510 PAL cycles on the diverse fixture. This does not meet the normal-C64 performance targets.

Qualification

The canonical build passes the five editor/cache cases, the exact public insert/delete sequence in Plain and Full at both widths, the existing wrapped and prewrite-fault suites, and the unchanged 68-byte staged stack bound. The paired private proofs also retain CRC, foreign-owner, stale-capability, mixed-helper-generation, NUL and source positions above64KiB coverage.

Native qualification includes20 typing/fallback frames,16 Plain/Full caret frames,11 source/settings/preview frames and6 controlled lifecycle frames. All53 were directly reviewed and independently checked against VIC pixels. All20 corrected typing screenshots are byte-identical to the previous control. Lifecycle testing includes resumed editing, Undo, SaveAs and reopen through the existing standalone adapter; it is not ReadyOS suspend/resume integration.

The59-overlay integration repeats the20 typing,16 caret and11 presentation frames; all47 were inspected. The20 typing screenshots remain byte-identical to the earlier54-overlay control. Additional tests cover partially visible GEOS carets, native lexical reuse, the phase/index regression, all256 CRC table words, 132 packet checks and a70,002-byte document on a1MiB REU. The unchanged68-byte stack bound includes the new finite lexical dispatch tables and admitted R25 capture calls. None of these checks closes the remaining structural-edit cost.

Normal-role80-column raster

R16 now proves an all-normal effective-role row before skipping tile color arbitration. Original glyph/flag conversion and mixed-role arbitration remain. NativePAL short insert/BACK improve about11.4% with identical pixels and the same four overlay loads; the separate wrapped shift route is unchanged. No resident/BSS/ZP/font growth occurs; R25/C33 are rebuilt against the actual new R16 CRC. See the proof and limitations.

Empty-CR physical-row insertion/removal

Plain60/80 ordinary insertion/BACK can now move a validated body suffix when wrapping adds/removes an empty CR physical row. Existing R24 operations preserve source/header/packet/provider authority, complete packets/anchors/CRbits and final R25 publication. NativePAL cases improve43.9–57.5% with identical images. Unknown offscreen exposure still falls back. Return-key split/merge now has the separate checked R29 path described below. See the geometry contract and measured limits.

Return at a short line's end

The long-row validator now authenticates an inserted byte before testing the old row's terminal bound. Previously a Return exactly at that bound could be mistaken for the original CR: source bytes were correct, but the following rows stayed at their old screen positions. The terminal CR now reaches the existing structural-edit admission; ordinary terminal glyphs still use the validated path. Deletion checks its old-byte bound before reading. No CRC, provider or source guard is removed. See the isolated regression.

Native-row input setup

C33 initializes the complete19-byte native-row input descriptor from a constant template, then sets the current width/count and physical row. It leaves the renderer-owned49-byte scratch area to the worker. This recovers40 code/data bytes without extra overlay calls or changing either native renderer. Both widths pass544 glyph/role/flag vectors (1088 calls) with zero versus nonzero scratch, identical independent pixels/attributes and instruction counts. Registered edit/fault, stack and native60/80 typing checks also pass. This is space recovery for further rendering work, not a structural-edit speed claim.

Plain Return splits and BACK merges

R29 now validates the complete visible logical paragraph against the genuine pending source, stages candidate rows, and paints only rows whose full240-byte packets differ. It copies/authenticates the real narrow font once and uses a local role0 blitter, reducing loads9→4. R25 still performs the final source/ revision/CRC VALID-last commit. Selected, folded, styled, panned and incomplete offscreen scopes retain the existing renderer. See the phase, authority and build contract and paired native timings.