docs: weekly backlog re-verification and MODULE_AUDIT refresh (August 23, 2026)

[cursor]

Summary

• Re-verified every open item in all 13 TODOs.md files against current source, re-deriving each citation. One item closed (the Locations ranking-reorder animation, by PR feat(WhereUI): animate live location ranking overtakes #289) against last week's zero.
• Filed four new items, three of them about the retained tooling this window landed: the new direct test suites fail on Linux instead of skipping; sync-agents needs a bash launcher; Tools/'s two docs are linked from nothing; and every Flyover registration titles its screen with an English literal.
• Corrected two published claims and four stale citations — including a Calendar.current recount that had drifted to 15 by folding in out-of-scope test files, and one item whose premise (not just its line numbers) was wrong about how DataIssueScanner gets its primary regions.
• Fixed two false Linux claims in the root AGENTS.md and the todo-triage skill, both invalidated this window when sync-agents gained a #!/usr/bin/env ruby shebang.
• Rewrote MODULE_AUDIT.md from what the backlog now says.

Why

The scheduled weekly pass of the todo-triage procedure. INBOX.md was empty, so this pass is re-verification, new-surface review, and the audit refresh.

Seven of the window's eight commits touched no shipping Swift — four hardening the retained Python/Ruby behind the root commands (#283, #284, #287, #288) and one rebuilding CircleCI's iOS side into build-once/shard-the-snapshots (#276). That changed what an audit on Linux can do: those suites need no Xcode, no Swift toolchain, no simulator, and no network, so they are the first tests here a Linux host can genuinely run. This is the first edition of this report that executed something rather than only reading.

Running them is what produced the pass's main finding. 61 of 62 Python cases and 62 of 75 Ruby cases pass on Linux; all 14 failures are platform assumptions rather than regressions, in three separable groups (macOS-only /usr/bin/lockf and /usr/bin/plutil, a pristine-HOME test that inadvertently needs a system interpreter, and one assertion pinning a shell's 126-vs-127 convention). Tools/ADVERSARIAL_TEST_PLAN.md already assigns each command a platform — the tests just don't read it, and because both CI systems run them only on macOS, nothing would have said so.

The second finding came from trying to run ./sync-agents as the skill instructs. PR #283 gave it a Ruby shebang, making it the only root command of eighteen that reaches its interpreter that way, so a bare call exits 127. The root AGENTS.md Linux table listed that exact invocation as working, and the todo-triage skill repeated it — so the automation's own instructions were wrong about the automation's own environment. Both are corrected here; the launcher is filed.

Review focus

• The closed entry (Where/TODOs.md, completed P2s). PR feat(WhereUI): animate live location ranking overtakes #289 shipped a different design from the one the item proposed: not .animation(_:value:) on the ForEach, but a layout that interpolates card positions while holding semantic order fixed, because moving the source hierarchy reorders VoiceOver mid-transition. The entry records that difference deliberately, so a future reader doesn't wonder why the one-line fix was skipped. Worth a check that it describes what actually landed.
• The two new root P1s — whether the platform-skip framing is the right shape, and whether a bash launcher is the fix you want for sync-agents versus documenting the prefix.
• The Calendar.current counting tripwire. A recount reached 15 this week by including five spelled-out uses under Where/*/Tests/, which are outside Bumper's production-only scope. The number is 12, unchanged for a third pass, and the item now carries the scoping rule so a sixth pass doesn't re-litigate it.
• Corrected premises, not just citations. The RegionDays/RegionRanking item claimed DataIssueScanner received primaryRegions from WhereSession; it derives them itself. Second week running that an item needed its reasoning rebuilt rather than its line numbers bumped — flagged as a theme in the report.
• The report's Limitations section now separates what was executed and passing from what was executed with failures analyzed from what remains unread, instead of the blanket "static analysis only" of prior editions.

⚠️ Breaking changes

None. Documentation only — no source, manifest, or CI changes.

Compatibility

No effect on build, targets, or runtime behavior. The AGENTS.md edit changes one row of the Linux checks table; ./sync-agents regenerated the CLAUDE.md mirrors, which are gitignored.

Testing

• ./swiftformat --lint — 0 of 1110 files require formatting.
• ./attribution --check — up to date, 12 credits.
• ./shellcheck — clean.
• ./snapshot-shards check — plan valid: 47 suites across 3 planned shards plus intake (1=13, 2=15, 3=18, 4=1). Confirms PR feat(WhereUI): animate live location ranking overtakes #289's new snapshot suite lands on the intake shard, which is what the documented design says should happen.
• mise exec -- ./sync-agents — regenerates cleanly.
• Direct tooling suites — run and their failures diagnosed (that analysis is one of this pass's findings, not an incidental red).
• Not run: anything needing a Swift toolchain or Xcode, so no ./test, no --architecture-only, no ./xcstrings --lint. The report says so in its Limitations, and no claim here depends on having run them.