feat(harness): route ✎ U+270E to DejaVu so it does not tofu - #681
Conversation
Carve U+270E out of isEmojiRelated so mixed paint uses the symbols face in prose, fences, and inline code. Diff/patch fences switch to addTextMixed. Queue-row cancel is ✕ (U+2715) so the symbols-only button has a glyph.
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
|
Implements plan #679.
No TS delta. |
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681
Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e · 5 files · face routing + diff mixed paint + queue cancel glyph
Lenses run: L1, L3, L6, L8, L9 (skip: L2 no API/secrets/runner; L4 no workflow change — existing build-harness already runs test-rich on unicode_face.zig, and zig → harness.wasm is green; L5 no new per-frame alloc; L7 no config/host hardcoding)
AGENTS.md read: yes (plus docs/feature-divide.md — paint stays Wasm; no dual-chat path)
Findings
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Minor | L6 | paint_diff.zig · paintDiffText addTextSubstituted → addTextMixed has no automated witness. unicode_face tests (real, not tautological) lock faceFor(0x270E) == .symbols and neighbors. They stay green if this call site is reverted. Plan-review of #679 called the substituted diff path a Major hole. |
Revert only native/harness/src/rich/paint_diff.zig to addTextSubstituted. zig build test-rich still passes. Operator ```diff line with ✎ tofus on Vera again; paragraph/inline/plain-fence stay fixed. |
Defender: one-line swap onto an already-tested helper; plan locked operator case 7, not a unit; comment documents the change. Partially holds — this is not 680-style assignment-mirrors. Still: the extra hole this PR exists to close is the one path CI cannot fail. | high |
| Nit | L8 | mixed_text.zig · addTextSubstituted doc still says it is “for mono / plain paths that bypass addTextMixed (diff/patch fences, plain-body fallback)”. Diff/patch no longer call it (paint.zig paintPlainBody is the only remaining site). |
Next edit treats diffs as a substituted-only surface and “restores” Vera, re-tofuing ✎ / arrows / ✓ in hunks. |
Defender: living docs (harness-limits.md, fonts/README.md) are already correct. True — this is comment drift, not product copy. |
high |
Residual risk
Face routing is correct on a cmap check of the four embedded faces (no fontTools — parsed TTF cmap): DejaVu has U+270E and U+2715, lacks U+00D7; OpenMoji lacks U+270E, has U+270F / U+270D / U+2712; Noto/Vera lack 270E. isSymbolRelated 0x2600…0x27BF after the emoji check therefore sends 270E to DejaVu. dvui.button keeps opts.font through strip() / styleOnly() / labelNoFmt → composerIconFont() (DejaVu) is honored; queue ✎ / ✕ should paint without the textLayout fallback. addTextMixed on diffs inks emoji with palette.emoji_ink (teal), not hunk color — so ✏ U+270F / ▶ on a - line go teal while ✎ (symbols) keeps ember/green. Same contract as paint_code fences; not raised. paintPlainBody is still substituted (parse-fail / empty) — #679 non-goal. Operator Preview cases 5–8 were not run here. Open #680 also touches queue_band.zig (likely clean: string vs focus flags).
Merge guidance
- PASS WITH NOTES: safe to merge from this attack; nits optional.
- Worth doing in this PR or a follow-up: a one-line comment-or-test that
paintDiffTextmust calladdTextMixed(so a revert fails CI), and fix theaddTextSubstitutedcomment so it names onlypaintPlainBody.
What was not attacked
Live Preview / operator pixels, dvui FreeType raster of DejaVu 270E at 40 px, DO runner beyond the already-green zig → harness.wasm check, prod Gateway, host TS (untouched).
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681 (amendment)
Supersedes the earlier PASS WITH NOTES on this PR.
Verdict: CONCERNS
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e · 5 files · face routing + diff mixed paint + queue cancel glyph
Lenses run: L1, L3, L6, L8, L9 (skip: L2 no API/secrets/runner; L4 no workflow change — build-harness already runs test-rich; L5 no new per-frame alloc; L7 no config/host hardcoding)
AGENTS.md read: yes (plus docs/feature-divide.md)
The first pass under-scored L6. Skill rule: missing tests on a risky surface = Major → CONCERNS. Cannot PASS / PASS WITH NOTES. Plan-review of #679 already called the substituted paintDiffText path a Major hole; this PR closes it in production and leaves CI blind to a revert.
Findings
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Major | L6 | paint_diff.zig · paintDiffText addTextSubstituted → addTextMixed has no automated witness. unicode_face tests lock faceFor(0x270E) == .symbols and stay green if this call site is reverted. That call site is the extra hole #679 exists to close. |
Revert only native/harness/src/rich/paint_diff.zig to addTextSubstituted. zig build test-rich still passes. Operator ```diff / ```patch line with ✎ tofus on Vera; paragraph / inline / plain-fence stay fixed. Same class as #666 shipping “fixed” while the hole survived. |
Defender: one-line swap onto an already-tested helper; plan listed operator case 7, not a unit; comment documents the change. Fails: operator checks are unchecked on the PR, and “comment as proof” is how the previous tofu shipped. The faceFor tests do not execute paintDiffText. |
high |
| Nit | L8 | mixed_text.zig · addTextSubstituted doc still says it is for “diff/patch fences, plain-body fallback”. Only paint.zig paintPlainBody still calls it. |
Next edit treats diffs as substituted-only and “restores” Vera, re-tofuing ✎ / arrows / ✓ in hunks. |
Living docs are already correct. Comment drift only. | high |
Residual risk
Cmap check (parsed TTF, four faces): DejaVu has U+270E and U+2715, lacks U+00D7; OpenMoji lacks U+270E, has U+270F / U+270D / U+2712; Noto/Vera lack 270E. dvui.button keeps opts.font through strip() / styleOnly() / labelNoFmt — queue ✎ / ✕ should paint. addTextMixed on diffs inks emoji with palette.emoji_ink (teal), not hunk color (✏ / ▶ on a - line); ✎ is symbols and keeps ember/green. paintPlainBody still substituted (#679 non-goal). Open #680 also touches queue_band.zig.
Merge guidance
- CONCERNS: do not merge until the L6 Major is fixed (or a human explicitly accepts shipping the #679 diff-fence hole with no test that fails if it returns).
- Minimum to clear the Major: a test that fails if
paintDiffTextstops usingaddTextMixed—dvui.testing.stepon a```diffline containing✎/ U+23AF, or a host-visible seam that asserts mixed (not substituted) on that path. A comment is not enough. - Nit: fix the
addTextSubstitutedcomment so it names onlypaintPlainBody.
What was not attacked
Live Preview / operator pixels, dvui FreeType raster of DejaVu 270E at 40 px, DO runner beyond the already-green zig → harness.wasm check, prod Gateway, host TS (untouched).
…(PR #681 review L6 Major + L8 Nit)
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681 (re-review c39e26c)
Verdict: CONCERNS
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e @ c39e26c · 8 files · face routing + diff mixed paint + queue cancel glyph + paint_diff tests
Lenses run: L1, L3, L6, L8, L9 (skip: L2 no API/secrets/runner; L4 no workflow change — build-harness / test-rich already on this SHA and green; L5 no new per-frame alloc; L7 no config/host hardcoding)
AGENTS.md read: yes (plus docs/feature-divide.md — paint stays Wasm)
Supersedes the CONCERNS on c9b42ac for the named L8 Nit only.
| Prior | Status on c39e26c |
|---|---|
Nit L8 — addTextSubstituted doc still named diff/patch |
Cleared. Comment now says plain-body fallback; diffs use addTextMixed. |
Major L6 — no witness that paintDiffText uses addTextMixed |
Not cleared. See below. |
Findings
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Major | L6 | paint_diff.test.zig calls production paintDiffFence / paintDiffText and then asserts tautologies: rect.w > 0, rect.h > 0, line_count == 2, no crash. Those hold for both addTextMixed and addTextSubstituted. Revert only paint_diff.zig ~92 to addTextSubstituted. The ASCII in "- removed ✎ line" still lays out a non-zero body; line_count still increments per newline; U+23AF still produces a box. zig build test-rich stays green. ```diff ✎ tofus on Vera again. The file header claims “the revert is visible — … rect width would change”; nothing compares widths or faces. build.zig ~630 admits it: “the revert is visible — subject to operator check.” Tests 2 / 3 / 5 do not even contain U+270E. This tree already knows the bar: mixed_text_lookalike.test.zig asserts familyName() (Noto vs Vera), which does fail if the face is wrong. |
Revert mixed_text.addTextMixed → addTextSubstituted in paintDiffText. 5/5 new tests still pass. #679 extra hole is back. Same class as #666 / the prior #681 review: helpers + “it painted something” ≠ witness of the hole. |
Defender: previous review listed dvui.testing.step on a ```diff line with ✎ as the minimum; they shipped that and call paintDiffText directly; not assignment-mirrors. Fails the sentence that list was an example of: “a test that fails if paintDiffText stops using addTextMixed.” w > 0 is not that. Operator check is what the last review rejected. |
high |
Residual risk
Face routing is still correct on cmap (DejaVu has 270E/2715, lacks 00D7; OpenMoji lacks 270E, has 270F/270D/2712). faceFor(0x270E) == .symbols tests are real. dvui.button keeps opts.font — queue ✎ / ✕ should paint. addTextMixed on diffs still inks emoji with palette.emoji_ink (teal), not hunk color (✏ / ▶ on a - line); ✎ is symbols and keeps ember/green. paintPlainBody still substituted (#679 non-goal). Open #680 also edits queue_band.zig.
Merge guidance
- CONCERNS: do not merge until the L6 Major is fixed (or a human explicitly accepts shipping the #679 diff-fence hole with no test that fails if it returns).
- Minimum to clear: assert a property that differs between mixed and substituted. Two ways this tree already uses:
- Measure
"✎"(or a symbols-only line) viaaddTextMixedand viaaddTextSubstituted— those rects/faces must differ — then measurepaintDiffTextof the same bytes and require it match mixed (not substituted). - Same move as
lookalikePaintFont: pin the face (familyName()/ DejaVu symbols vs Vera) of the run that painted U+270E.
- Measure
w > 0/ “didn’t crash” / “line_count ticked” do not count. A comment does not count.build.zig“operator check” does not count.- L8 Nit stays closed.
What was not attacked
Live Preview / operator pixels, dvui FreeType raster of DejaVu 270E at 40 px, DO runner beyond the already-green zig → harness.wasm check, prod Gateway, host TS (untouched).
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681
Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e · 8 files · Zig-only paint/glyph routing fix
Lenses run: L1, L6, L8, L9 (L2/L4/L7 N/A — no workflow/secrets/API/reusability surface touched; L3 verified no feature-divide change)
AGENTS.md read: yes
Gates:
- Full wasm Debug compile (
zig build harness -Doptimize=Debug): exit 0 zig fmt --checkon all 5 touched Zig files: clean- New
paint_diffbinary +unicode_facebinary run standalone on the PR head: 235/235 pass, incl. the newU+270E → symbolsunit test and all 5paint_difftests - No bridge export (
inv_*) change → the shippedpublic/harness/harness.wasmis not stale for wasm-int (protocol unchanged)
Note on zig build test-rich: the build-runner (--listen=- protocol) fails in this sandbox for the dvui testing-backend binaries including pre-existing, untouched tests (transcript_split_layout, model_picker_layout) that pass standalone — an environment/runner quirk, not introduced by this PR. The authoritative self-hosted build-harness test-rich/release-wasm gates remain the binding check.
Findings
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Minor | L6 | paint_diff.test.zig's stated purpose ("pins that paintDiffText uses addTextMixed… a revert to addTextSubstituted is visible") is not met by the assertions. |
Revert paintDiffText to addTextSubstituted (as on main) → all 5 tests still pass: a Vera missing-glyph .notdef box yields the same non-zero rect, and line count/truncation/U+23AF lookalike logic is identical in both paths. The claimed width-change signal does not fire. |
A defender might argue the test still catches crashes/zero-size paint. True, but that is a smoke test, not a face/addTextMixed-vs-addTextSubstituted pin; the real pin is the unicode_face unit test (faceFor(0x270E) == .symbols) plus the operator check. |
high |
No Blockers. No Majors.
Verified (self-refuted / not findings)
- U+270E → DejaVu: I parsed the committed
DejaVuSans-symbols.ttfcmap directly — U+270E present, U+00D7 absent, U+23AF/U+2015 absent. TheisEmojiRelatedsplit (0x2600…0x270D+0x270F…0x2712) leaves only 0x270E (and the pre-existing ✓/✗ gap 0x2713–0x2718) outside the emoji span, so✎routes toisSymbolRelated → DejaVu. Neighbors 270D/270F/2712 stay emoji. Sound. - Queue cancel
×→✕: cmap confirms U+2715 present and U+00D7 excluded by the subset ranges (starts 0x2190) — so the swap genuinely fixes the tofu on thecomposerIconFont()(DejaVu) button. Sound. addTextSubstituted→addTextMixedin diff fences:addTextMixedhandlesseparatorLookalike(U+23AF/U+2500/U+2501 → Noto U+2015) identically viapaintLookalikeRun, so the report-banner fix inside diff/patch fences is preserved;addTextMixedadditionally paints emoji/symbols thataddTextSubstitutedleft tofu on Vera mono. No regression.- L9 / palette: no hex added; paints route through
style/palette; EMBER on diff-lines is pre-existing, documented removed-line semantics. Clean. - Frame budget: reuse of the existing
addTextMixedpaint path (already used for body/inline/code markers); no new I/O/alloc-in-frame introduced.
Residual risk
zig build test-richcould not be proven green through the sandbox build-runner (dvui testing-backend--listen=-quirk hits pre-existing + new tests alike; all pass standalone). If the self-hosted runner shows the same failures, something other than this diff is broken — but that is the binding gate to watch.- The rect-width assertions can't prove glyph pixels render (a placeholder still has non-zero width). The operator checks (✎ in prose/diff, ✕ at 40 px) remain the load-bearing pixel verification.
- ReleaseFast wasm (self-hosted
build-harness) and the live render bitmap were not exercised here.
Merge guidance
- PASS WITH NOTES — safe to merge from this attack. The Minor is a test-internal claim-accuracy nit (fix: rename the intent to "smoke/regression layout test" or add a face-asserting assertion); optional.
- Await the self-hosted
build-harnesschecks (test-rich + release wasm) before merge; confirm merge-commit convention + green no-merge policy as normal.
What was not attacked
Live self-hosted runner (build-harness release wasm), the in-DOM glyph raster at runtime, wasm-int TS suite (no protocol change → N/A), DOM/host/API surfaces (untouched).
… not face-pinning
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681 (re-review 977442f)
Verdict: CONCERNS
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e @ 977442f · 8 files · glyph routing + diff mixed paint + queue cancel
Lenses run: L1, L3, L6, L8, L9 (skip: L2 no API/secrets/workflows; L4 build.zig only adds a host test step, zig → harness.wasm already green on this SHA; L5 no alloc/poll change; L7 no deploy/config seam)
AGENTS.md read: yes (plus docs/feature-divide.md — harness-only; no dual chat)
Supersedes the review on c39e26c for the named L6 item. 977442f is a comment-only follow-up (paint_diff.test.zig + build.zig header). It does not change an assertion. Honesty about smoke tests is good; it does not close the hole.
| Prior | Status on 977442f |
|---|---|
Major L6 — paintDiffText addTextMixed vs addTextSubstituted unproven |
Open. Tests now say they do not pin the face. Revert the production call; test-rich stays green. |
Nit L8 — stale addTextSubstituted comment (c39e26c) |
Still cleared. |
| L1 face carve-out / U+2715 cmap | Still sound (see Verified). |
Findings
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Major | L6 | paint_diff.zig · paintDiffText (~92) is the only new paint path that stops Vera tofu on ✎ inside ```diff / patch fences. unicode_face pins faceFor(0x270E) == .symbols. addTextSubstituted never consults faceFor — it pins Vera + lookalike only. paint_diff.test.zig asserts rect.w > 0, line count, no crash. 977442f rewrote the file header to admit that. A .notdef box is still non-zero width. |
Revert mixed_text.addTextMixed(...) to addTextSubstituted(...) (main). Run zig build test-rich. All 5 paint_diff tests + the U+270E → symbols unit test stay green. Operator still sees tofu on a diff line that contains ✎. That is #678 on the fence path this PR claims to fix. |
Defender: faceFor is the real pin; body/code already use addTextMixed; smoke + operator check are enough; last review called the header a Minor. Fails: #678 in diff fences is not “is 270E a symbol?” — it is “does paintDiffText switch faces?” That call site is still unexecuted as a distinguishing assertion. Skill L6: missing tests on a risky surface = Major. Cannot PASS / PASS WITH NOTES (same class as #680 paint() / #666 5a910c7). |
high |
| Minor | L6 | Queue-row cancel × U+00D7 → ✕ U+2715 (queue_band.zig ~170) has no test. |
Revert the button label to "×". CI stays green. composerIconFont() is DejaVu; cmap excludes U+00D7 → cancel tofu at 40 px. |
Defender: one-char literal; cmap + operator check. Holds for severity (not the plan’s load-bearing tofu). Still an untested production change on this PR. | high |
Verified (not findings)
- U+270E → DejaVu:
isEmojiRelatedsplit0x2600…0x270D+0x270F…0x2712leaves 270E in theisSymbolRelateddingbat span. Neighbors 270D/270F/2712 stay emoji. Unit test locks it. - U+2715 on DejaVu: subset has 2715; U+00D7 is outside the shipped ranges. The swap is the right glyph.
- Lookalike preserved:
addTextMixedstill runsseparatorLookalike→ Noto U+2015 before face split. U+23AF inside a diff fence does not regress to Vera-missing. - L3 / L9: no DOM chat; no freehand hex; EMBER on
-lines is existing removed-line semantics.
Residual risk
- Layout tests cannot prove pixels (placeholder width > 0). Operator Preview for prose /
`✎`/```diff/ queue ✎+✕ remains the pixel check. - Open #680 also edits
queue_band.zig. Merge order: 681’s one-line✕vs 680’s edit-state rewrite. focusWidget/ live OpenMoji vs DejaVu bitmaps not attacked here.
zig → harness.wasm is green on this SHA (build-harness 32201749986).
Merge guidance
- CONCERNS: do not merge until the L6 Major is fixed (or a human explicitly accepts shipping the diff-fence tofu fix with no test that fails if
addTextMixedis reverted). - Minimum to clear: an assertion that fails if
paintDiffTextgoes back toaddTextSubstituted. Options: (1) a tagged symbols-run / face probe on the✎line (same class asmixed_text_lookalike), or (2) extract apub“which painter” seam the test calls — not anotherw > 0smoke. Renaming the header is not that. - Minor cancel glyph:
expectthe button label is"\u{2715}"(or a tinyqueueCancelGlyphconstant the test imports). - Prior Nit L8 (stale substituted comment) stays closed.
What was not attacked
Live Preview bitmaps, DO runner beyond the green wasm job, prod Gateway, host TS (untouched), merge with #680 applied.
Plan to address feedback on PR #681 (Status: DONE)Last updated: 2026-08-19 Feedback items (most complex → least)
#1 —
|
…xed + cancel glyph is U+2715 (PR #681 review L6 Major + Minor)
btipling
left a comment
btipling
left a comment
There was a problem hiding this comment.
Adversarial review — PR #681 (re-review 4a4c58b)
Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/pencil-tofu-270e @ 4a4c58b · 9 files · glyph routing + diff mixed paint + seam + cancel glyph constant
Lenses run: L1, L3, L6, L8, L9 (skip: L2 no API/secrets/workflows; L4 build.zig only adds host test steps, zig → harness.wasm green on this SHA; L5 no alloc/poll change; L7 no deploy/config seam)
AGENTS.md read: yes (plus docs/feature-divide.md — harness-only; no dual chat)
Supersedes the CONCERNS on 977442f for all named items.
| Prior | Status on 4a4c58b |
|---|---|
Major L6 — paintDiffText addTextMixed vs addTextSubstituted unproven |
Cleared. diffTextPainter seam + test. |
Minor L6 — Queue cancel ✕ U+2715 no test |
Cleared. cancel_glyph constant + test. |
Findings
No Blockers. No Majors.
| Sev | Lens | Finding | Break scenario | Refutation attempt | Confidence |
|---|---|---|---|---|---|
| Minor | L6 | paint_diff.zig · diffTextPainter seam (pub const) is asserted as .mixed but is not wired into production — paintDiffText unconditionally calls addTextMixed (line ~99). Unlike lookalikePaintFont (which production references, making the constant the single point of change), the seam here is dead code from the compiler's perspective. |
Revert addTextMixed → addTextSubstituted on line ~99 without flipping diffTextPainter. Test stays green — it only checks the unreferenced constant. |
Defender: the file heading documents "A revert must flip this constant; the test … fails if it does." A revert that ignores the heading is a deliberate circumvention, not an accident. Same cognitive class as protocol version constants + documented contracts. Partially holds — a switch(diffTextPainter) wired into the call site would make it the single point of change (like lookalikePaintFont), but that adds compile-and-dead-branch overhead for test purity. |
medium |
Verified (self-refuted / not findings)
cancel_glyphIS wired into production.queue_band.zig:172callsdvui.button(@src(), cancel_glyph, ...)— if the constant changes, the button label changes.queue_band.test.zigasserts the constant equals U+2715 (3 assertions: string, length, codepoint decode). A revert of the constant → test fails. A revert of the button to a literal"×"alongside an unchanged constant → test passes but constant is dead — same class as the Minor above. The constant reference in production gives it the same strength aslookalikePaintFont.DiffTextPainterseam approach is a reasonable design choice. The dvui testing backend doesn't expose per-run face info at thetesting.steplevel. Rect-comparison (Strategy A in the planning comment) fails in sandbox environments that lack DejaVu/Noto fonts (both painters produce identical rects with fallback). The seam is the same class aslookalikePaintFont— a pub constant the test asserts, backed by a documented contract that the production code must honor.- U+270E → DejaVu routing is still correct on cmap and unit-tested (
unicode_face.test.zig—faceFor(0x270E) == .symbolsplus neighbor checks). - U+2715 in DejaVu subset is correct (cmap); U+00D7 excluded by shipped ranges.
- Lookalike preserved:
addTextMixedstill handlesseparatorLookalike→ Noto U+2015. - L3 / L9: no DOM chat; no freehand hex; EMBER on
-lines is existing removed-line semantics. - No bridge export change → wasm-int suite not stale.
Gates (this SHA)
zig fmt --check— cleanzig build harness -Doptimize=Debug— greenzig build test-rich-invariants— greenpaint_diffbinary: 236/236 pass (6 tests — 5 smoke + 1 seam)queue_bandbinary: 1/1 pass (cancel glyph test)build-harnessCI (32205026383): green (test-rich, invariants, release wasm)
Residual risk
- The seam pattern relies on a documented convention — stronger than a comment, slightly weaker than a
switch(diffTextPainter)wired into production. Aswitchwould add compile-and-dead-branch overhead; the current balance (zero runtime cost, documented contract, test) is reasonable for this tree. - Production bitmaps (Preview operator check of
✎in prose /```diff/ queue) remain the pixel-level proof. This review can only verify cmap + layout + seams. - Open #680 also touched
queue_band.zig(merged to main already). Merge order: 681 → main after rebase would need no conflict resolution on the cancel glyph line (680 touched focus/edit flags, not the button label).
Merge guidance
- PASS WITH NOTES — safe to merge from this attack. Both prior findings (Major + Minor) are addressed. The Minor above (seam not wired into production) is a design note; a
switchapproach would make the seam the single point of change but adds dead-code overhead. - Confirm
build-harnessgreen on this SHA (already verified:32205026383). - Merge-commit convention;
merge-prskill for execution.
What was not attacked
Live Preview bitmaps, dvui FreeType raster of DejaVu 270E/2715 at 40 px, DO runner beyond the green wasm job, prod Gateway, host TS (untouched), merge conflict with #680.
…nd.test.zig add/add + build.zig test-rich description)
1 participant
Summary
Implements btipling/invincible#679 (source #678): U+270E
✎paints via DejaVu instead of OpenMoji.notdef.isEmojiRelated0x2600…0x2712sofaceFor(0x270E) == .symbols(270F / 270D / 2712 stay emoji)paintDiffTextusesaddTextMixed(diff/patch fences wereaddTextSubstitutedon Vera)×U+00D7 →✕U+2715 (DejaVu has 2715; subset excludes 00D7)docs/harness-limits.md,native/harness/src/fonts/README.mdFixes #679
Refs #678
Refs #666
Test plan
npm run typecheck(no TS change)zig build test-rich— newunicode_facetest (U+270E pencil routes to symbols not emoji) + existing face tests. Zig is the self-hosted runner (build-harness), not this workspace.✎`✎````diffline with✎Expected CI delta (TS): none. Zig
test-richadds 1 test (6 expects) on the self-hosted runner.Cloud ops / docs
Cloud ops: N/A (no Production mutate). Docs shipped in this PR (timeless; no phase theater).