experiment(core): layered paint grid with measured performance tradeoffs

[kommander]

Status

Draft opt-in candidate. Explicit startup-cost decision required; the universal
performance goal remains unmet. The owner permits one or two1-3% benchmark
losses with substantial remaining gains. Minor warm losses fit that tolerance;
larger first-use/mixed costs do not become acceptable by averaging them away.

Runtime remains d86dd308, secured at a8801276, versus actual main 202e1e6a.
The focused continuation retained no runtime change. Recommendation: keep this
low-complexity candidate disabled by default and explicitly accept the disclosed
startup cost for opt-in selective scenes, or decline it. Do not automatically add
another recording strategy or mark the PR ready. This is a tradeoff recommendation,
not a proof that no future architecture could improve startup.

Architecture

• Target-owned glyph/style spans, bulk fills and ASCII TextBuffer/drawText inputs;
  scalar handling remains for wide/complex/shared-tab inheritance semantics.
• One command list and owned arena, old/new streams for comparison, bounded
  wide/inherited dependency closure. No reverse cell index or duplicate renderer.
• Initial/consecutive known-full frames skip recording transactions and initial
  grid allocation. Native owns invalidation/recovery; the JS flag owns only the
  outstanding recording obligation. Raw writes, late fallback and scope cleanup
  remain intact. Root uses its existing geometry revision, not duplicate state.
• Whole runtime diff:10files,+1135/-72, including623line engine and+260/-67 native
  buffer integration. Tests/reports excluded from those counts, not removed.
  Redundant owner fields and identical raw/direct modes were already removed.
• New direct-binding trial removed forwarding methods with no native ABI change,
  preserving status2 throw BEFOREtry/finally. It was measured and rejected, not
  shipped for fractional warm noise. No new runtime memory or strategy framework.

Fresh First-Use Evidence

• Two five-repeat actual-main/current focused batches, local/layout/generic,
  120x40 and240x80 bothdepth4, ON/OFF,1000/600 steady frames after all28mixed frames.
  Same observer-free harness, rotating sources, verified normal binaries. No
  forcedGC, dropped tails, hidden prewarming or inferred pause subtraction.
• First capture remains0.94-1.20ms small versus0.38-0.43ms ordinary main frame;
  large1.28-1.67ms versus0.63-0.75ms. These ranges are workload/batch medians, not
  per-repeat ranges. All first2/3 totals and individual distributions are linked.
• Smallgenericmixed+32.3/+21.3% across the two batches. Smalllocal+0.2/+4.4%,
  layout+1.8/+7.9%. Large mixed signs vary versus prior matrices; full ranges and
  pauses retained, no claim old large losses were GC or have universally vanished.
  -20fresh phase JSC profiles show Root first capture smallLLInt/largeBaseline,
  with scope and native text work also sampled. Not proof the entire gap isFFI/JIT.
• Seven alternating baseline/direct-binding repeats: capture1-11%better, but
  smallfirst3+3-6%, layoutmixed+6.2/+5.9%. Rejected and fully restored. Small
  constructor+first3 separately19.685->20.300ms; no constructor-hidden benefit.

Full Warm Matrix

Historical final simple-* source/binary-identical results, not newly rerun:
allnine ON/OFF, seven1000-frame repeats at120x40/depth1, five600 at240x80/depth4,
with rotating main/current/prior and separate original-driver observer family.

• Small selective warm means: unchanged-49.5%, localized-22.3%, outside-67.9%.
  Generic-11.1%, raw-12.4%, image-17.8%. Small losses:layout+2.7%, scroll+2.4%,
  broad+0.6%; layout/scrollp50 improve, broadp50+1.3%.
• Large selective means:-55.9/-32.6/-75.9%. Layout-7%, generic-5.2%, raw-9.1%,
  image-12.5%, scroll-4.5%; broad+2.1%,p50+0.4%.
• Historical mixed losses include smalllayout+17.4%,local+12%, large local+10.6%,
  generic+11.6%. These remain published, not replaced selectively by fresh wins.
• Original2016normal/transition four-channel matches remain exact-source evidence.
  Initial native grid allocation0; retained capacity about94/173KB small and
  243/542KB large plain/outside, notRSS or cumulative allocation traffic.

Validation

• Fresh restored Core17tests/260assertions; trial17tests passed too. No weakened
  coverage. Final docs-only fmt/lint/diff checks pass.
• All705baseline source/library pins and234measurement pinsets verified. Both
  normal libraries03acd5fa. Only the rejected trial'stwoTSfiles differed.
• Prior exact-source full rootReleaseFast/native2134/Core5500/Node26Core4757/
  packedBun+Node/React59/Solid271 remain evidence, not newly rerun counts.
  Earlier6890Solid crash and passing rerun stay disclosed; finald86Solid passed.
• Primary HEAD, dirtylock and index preserved exactly; no otherrefs/PRs, host,
  toolchain, newworktree/cache or cleanup changes. Same draft PR; no merge/ready.
• CI is observed once after this publication. No green-CI claim or polling.

Reports

• First-use tradeoff and whole-runtime explanation
• Fresh focused measurements and rejected trial
• All focused distributions, OFF, frames, sampled summary, patch and provenance
• Historical final full warm/original matrices
• Historical final full distributions and counters