v3.5.1.0 fix: repair the landing-page guard and the mutation harness that hid its holes

[savvides]

Summary

Three phone-only defects on idstack.org, and a repair of the test harness that was supposed to catch them.

The mutation suite could not fail. fresh() copies the repo into a throwaway directory per mutation and omitted extension/, so test-extension.sh failed on every copy before any mutation was applied. Because expect_fail only checks for a non-zero exit, all 36 mutations reported GUARDED regardless of whether their guard worked. The suite whose job is proving the other tests work had been unable to fail since the extension landed. Fixed, plus a null-mutation control that aborts when an unmutated copy is already red.

idstack.org on a phone

• .section had no safe-area inset, so every content region sat under the notch once viewport-fit=cover landed. The four containers now share one rule.
• Three ad-hoc touch floors (36/40/42px), with the footer controls applying theirs only below 480px — leaving tablets and desktop at 36.6px. One --tap-min token, 44px, everywhere.
• overflow-x: clip on html/body hid horizontal overflow instead of preventing it, turning a real overflow into unreachable content. Removed; the grid tracks that contain long install commands size with minmax(0, …).

New: the page is tested by rendering it. test/test-rendered-landing.js loads the page in headless Chrome and asserts the outcome at 11 widths. The existing suite checks CSS as text, which can only forbid spellings someone thought of — six kinds of edit shipped a broken page past it. All six fail the rendered suite:

Evasion	Text suite	Rendered
.copy-btn, .btn-badge { min-height: 30px }	PASS	fail
:is(.copy-btn) wrapper	PASS	fail
@container hoist of the floor	PASS	fail
@media only print and (max-width: 480px)	PASS	fail
@media (min-width:900px){@media(max-width:480px){…}}	PASS	fail
<style> inside an HTML comment	PASS	fail

Guard hardening. Assertions accepted the regressions they existed to catch: (clip|hidden) allowed the overflow-x: hidden that breaks the sticky nav; grid-template-columns: 1fr matched the 1fr 1fr it forbids; a --tap-min existence check passed while a later rule shipped 30px targets. Guards also no longer fail correct changes (adding a fifth container, reformatting minmax(0, 1fr)).

Test Coverage

63% (33/52 paths), 19 gaps, no regressions. Above the 60% minimum, below the 80% target — accepted, with the two largest gaps closed structurally by the rendered suite rather than by more text assertions.

Coverage was measured by mutating each parser helper branch-by-branch and re-running: a branch counts as covered only if breaking it breaks something.

Coverage diagram

[+] test/test-responsive-landing.js  (parser logic)
  ├── stylesheet(doc)        7 of 8 branches unguarded  [GAP]
  ├── mediaBlocks()          6 covered, 3 gaps
  ├── unconditional()        4 covered (self-test + mutation 26)
  ├── allMediaBodies()       entirely unguarded          [GAP]
  ├── present()              no-op mutant survives       [GAP]
  └── expect()               covered by all 9 mutations

[~] docs/index.html  (CSS invariants)   18/20 covered
  ├── notch-safe gutter, --tap-min uniqueness, root overflow (longhand +
  │   shorthand), unconditional floor, prelude scoping, 1fr collapse — all
  │   mutation-backed
  └── [GAP] rendered viewport overflow — NOW CLOSED by test-rendered-landing.js

SUBTOTALS: JS helper paths 13/26 (50%) | docs/index.html invariants 18/20 (90%)
COVERAGE: 33/52 (63%)  QUALITY: ★★★:14 ★★:18 ★:1  GAPS: 19

Pre-Landing Review

Six passes: 4 cleanup agents, 5 specialists, a coverage audit, a red team, and a Claude adversarial pass. 28 findings, all actioned. Every critical was in the test file, never in the page — the CSS came back clean from all three agents that measured it (container geometry byte-identical at 8 widths, zero overflow at 12 widths unclipped, type ratios preserved within 1%).

Specialists: performance NO FINDINGS · simplification 1 advisory · design 1 · maintainability 13 · testing 13.

Two findings were rejected on measurement, both claiming a removed min-width: 0 would overflow the page: reproduced as identical geometry, zero overflow. overflow-x: auto / overflow: hidden on the same element already zeroes the automatic minimum size.

Design Review

Design specialist: 1 finding (a wrong figure in DESIGN.md, fixed). Verified clean by measurement — container geometry unchanged before/after at 8 widths, no heading/body inversion from the collapsed clamps, focus ring keeps 8px headroom inside .install-pill, and the taller controls read as normal proportions.

Eval Results

No prompt-related files changed — evals skipped.

Adversarial Review

Claude adversarial pass: 3 blocking findings, all fixed and mutation-guarded. Codex did not run — the CLI is installed but its vendored binary is missing (spawn .../codex-darwin-arm64/vendor/.../codex ENOENT), so this PR has no cross-model coverage. npm install -g @openai/codex restores it.

Verification

Check	Result
CI, all 3 legs	362/362
CI mutation job	37 guarded, 0 unguarded
Rendered suite in CI	ran on all 3 legs (not skipped)
Layout, 11 widths	0 overflow, sticky nav, 44px targets

CI was red on the first three pushes: WebSocket is not defined. The rendered suite drives Chrome over the global WebSocket, which node exposes unflagged only from 22.4, and the setup-node step pinned 20. Invisible locally on node 24. Fixed in 4a5aafd.

Follow-ups (not in this PR)

• env(safe-area-inset-*) has zero executed coverage. The notch-safe gutter that motivated this refactor is asserted by declaration text alone. CDP cannot set safe-area insets; this needs a device or simulator.
• static.yml does not gate on test.yml. Pages deploys on push to main regardless of a red suite, which undercuts this entire regression gate. Pre-existing.
• check-doc-accuracy.py:152 pins the literal "353 assertions", two revisions dead, so the guard that should catch documented-count drift is unreachable. It silently permitted 353→355→361→362.
• check-doc-accuracy.py carries two byte-identical blocks for CLAUDE.md and CONTRIBUTING.md; this PR had to make the same (?:sh|py|js) edit twice.
• unconditional() and allMediaBodies() are exact complements computed by two separate @media walks; one 12-line split() returns both. Verified byte-identical, ~17 lines removable.
• docs/superpowers/plans/2026-08-20-…md still prescribes the three ad-hoc touch floors and the min-width: 0 this PR replaced. An agent told to implement it would reintroduce three mutation-guarded defects.