feat(harness): str_replace L2 TEAL/EMBER side bands

[btipling]

Summary

Implements #691 (source #687).

Successful str_replace L2 no longer dumps the whole audit block as one teal_text mono slab. Wasm paints:

• status line (chrome, teal_text)
• old on ember_surface / ember_text
• new on teal_surface / teal_text

-old_string / +new_string stay on the payload (emit + 20/6 preview) and are stripped from paint. Empty new still mounts a one-mono-line TEAL strip. Other tools keep the single-block path.

No protocol bump. No TS emit/preview change.

Files

• native/harness/src/rich/toolrun.zig — splitStrReplaceDetail (line-exact sentinels, slices into detail, no alloc) + 8 unit tests
• native/harness/src/ui/toolrun.zig — paintToolRun L2 side bands; id_extra +4..+8; min height = fontMono().lineHeight()
• docs/harness-limits.md — Tool-run L2 row: payload vs paint

Verification

Gate	Result
tsc --noEmit	green
vitest run --changed	no TS test delta (this PR is Zig + docs). Incidental wasm-int 9 fails here are missing harness.wasm in this sandbox — not this diff
zig build test-rich / test-rich-invariants / harness -Doptimize=Debug	CI only — no zig in this agent workspace
build-harness	required check after push
Caps	existing STR_REPLACE_DIFF_SIDE_MAX_BYTES = 4096 and 20/6 windows unchanged
Cloud ops	N/A
Living docs	in this PR

Expected vitest pass count: unchanged (0 new TS tests). New coverage is the 8 Zig splitter cases on build-harness / test-rich.

Test plan (operator, after Preview wasm)

• Expand a successful str_replace L2 → two bands, no visible sentinels
• Delete (new_string empty) → old band + visible empty TEAL strip
• exec / read_file L2 still one teal_text block
• ~390 px readable, no horizontal gutter
• Refresh: restored str_replace L2 still splits

Fixes #691
Refs #687

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 19, 2026 7:54pm

[btipling]

Implement-plan for #691.

• splitStrReplaceDetail (line-exact sentinels, no alloc) + 8 test-rich cases
• paintToolRun L2: EMBER old / TEAL new bands; empty new keeps one mono-line strip
• docs/harness-limits.md distinguishes payload sentinels vs paint
• tsc --noEmit green; no TS emit/preview change
• Zig compile + test-rich are the build-harness job (no zig in this workspace)

Not merged. Suggested next: adversarial-review.

[btipling]

Adversarial review — PR #692

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/str-replace-side-bg @ 53abb6f · 3 files · str_replace L2 EMBER/TEAL side bands (Wasm paint only)
Lenses run: L1, L3, L5, L6, L8, L9 (skip L2: no trust-boundary / emit / redact change. skip L4: no workflow edit; build-harness already ran zig build test-rich + wasm. skip L7: no deploy bind)
AGENTS.md read: yes · docs/feature-divide.md yes (paint stays in-canvas; no DOM chat)

Findings

No surviving Blocker / Major after self-refutation. Splitter traces match appendStrReplaceDiff + buildStrReplacePreview (line-exact sentinels, escaped -old_string / +new_string is content, … (N more lines) is content, ERROR / one-sentinel → null → existing single-block path). id_extra +4..+8 stay inside the 1024-wide item namespace and are mutually exclusive with the fallback +4 body. Palette uses ember_surface/ember_text and teal_surface/teal_text (not ember_accent). No protocol bump.

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Nit	L8	native/harness/src/ui/toolrun.zig still documents “up to 5 widget slots” and +4 detail body (~177–179). Paint now mounts +5 old box, +6 old tl, +7 new box, +8 new tl.	Next L2 widget added at +5 by following the comment collides with the old band (expander state / hit-target mix).	1024-wide stride still has room; a reader 20 lines down sees the new extras. Fails as documentation of the locked collision rule — plan-review #691 called unique extras a Minor for a reason.	high

Residual risk

Attack did not breach a runtime path with current evidence.

• Goal 2 (empty new → visible TEAL strip) rests on dvui honoring min_size_content.h = fontMono().lineHeight() on a box whose child textLayout received no glyphs (mixed_text.addTextMixed returns at len==0). 2 px pad alone was already rejected as insufficient. Splitter tests lock new.len==0; they do not paint. Operator checklist in the PR body is still unchecked.
• teal_surface (#0a1215) on transcript teal_bg is a small delta — empty strip can look like a hairline rather than a band at some displays.
• Fail-soft: malformed / host-fallback str_replace detail without both sentinels still paints one teal_text slab (headers visible). Correct, but a clipped preview that lost +new_string degrades silently.
• No 390 px wrap / no-h-gutter proof in CI.

Merge guidance

PASS WITH NOTES — safe to merge from this attack after the operator smoke (expand success, delete strip, exec/read_file still one block). The Nit is optional (+0..+8 in the slot comment).

zig → harness.wasm is green (test-rich includes src/rich/toolrun.zig). Do not treat this as --approve.

What was not attacked

Live Preview paint, restored-session L2 after refresh, Copy of the still-sentinel payload, dvui testing-backend layout of the empty new box, full vitest (N/A — no TS).