Clear unfilled stubs from the generated indexes and ignore coverage output

[Jason-Vaughan]

What

Janitor sweep — repository hygiene only. No source files, no tests, no behaviour change.

• FEATURES.md — removed ten auto-stubbed **TBD** — touched in this session: <path> entries under two dated headings. Each recorded only that a file had been touched, which git already records, and not one named a feature — the file's entire subject. Eight named a design document; the other two named PROJECT-MAP.md and CONTRIBUTING.md.
• PROJECT-MAP.md — the tanglebrain/ entry stopped mid-sentence and the docs/ entry was still the literal <!-- describe --> placeholder. Both now describe the tree as it is, leading with cli.py because project-state.yaml names it the primary surface.
• .gitignore — coverage output (.coverage, .coverage.*, htmlcov/) was not ignored.

Both index files are TangleClaw-maintained scaffolds whose auto-population is currently broken, with a fix expected upstream. The files and their structure stay deliberately — including the empty Server / API and Governance / Engines headings the wrap-step appends into. Only content that was never captured in the first place was removed; nothing was filled in by hand that a generator is meant to own.

Why

These are two of the first files a stranger opens, and this repo now has outside contributors. A map whose entry stops mid-sentence and an index whose visible content is ten TBD placeholders are worse than either file being absent — they read as abandoned rather than pending.

Separately: .coverage is a build artifact, and it had already reached a branch once.

Also in this sweep, not in this diff

• 18 stale local branches deleted. git branch --merged reported none of them — main squash-merges, so branch commits never appear in main and the built-in check is blind here. Mapping against the merged-PR list found 17 leftovers plus pr112. The four apparent orphans were checked for unique content before deletion: chore/oss-r0-gitignore's six gitignore patterns are all already in main, and the three wrap/* branches only re-added the same scaffold header. SHAs recorded locally in .prawduct/deleted-branches-2026-09-06.txt.
• First Norm Health sweep since ratification. Erosion measured at zero across five enforceable norms; recorded in project-state.yaml so the next sweep reads a trend rather than re-deriving a point. The broad-catch norm holds at 11 sites — 10 explicitly waived, the 11th re-raising and so exempt from BLE001 by construction, with no per-file ignore hiding anything.
• One norm trigger examined and re-affirmed unchanged. The formatting norm revisits "when a second regular contributor arrives and style starts costing review time". A contributor arrived; the cost did not — four contributor PRs reviewed this session, not one finding about style. It was first reported as fired, which was an overstatement of a conjunction. Owner re-affirmed the norm and declined to widen it.

Test plan

• make test — 664 passed, 17 skipped, ruff and mypy clean. (664 rather than 673 because this branch is off main and does not carry Note a lost measurement write, once per process #148's tests.)
• Line widths: only the lines this branch already touched were rewrapped. CHANGELOG.md carries ~90 pre-existing lines over the ~100-col soft wrap (one at 185); reflowing those would be the reformatting the no-formatter ruling exists to prevent.
• No source files touched, so the suite is a regression check rather than evidence for new behaviour.
• git check-ignore confirms the new patterns resolve, with no duplicate coverage entry.
• Verified nothing in the repo reads either removed FEATURES.md section, and that every docs/design/* file named in the new docs/ description exists.

Review

• Critic, round 1 rev-20260906T192656Z-2515796f — 0 blocking, 2 warnings, 3 notes. Both warnings were errors in my own prose: an unchecked "every one named a design document" that was false in 2 of 10 cases, and a map that omitted cli.py, the primary surface it exists to point at. The first landed in the same commit whose message cites the sibling rule about copying counts.
• Critic, round 2 rev-20260906T193251Z-8582657f — 0 blocking, 0 warnings, 6 notes. Three folded into the second commit (an unchecked universal in PROJECT-MAP.md — the same rule again, in the sibling file), three accepted.
• Independent PR review — 0 blocking, 0 warnings, 2 notes, both ridden along. It fact-checked every claim the diff adds against the tree, including scanning the reflog and unreachable objects to confirm .coverage really had reached a branch (it had: 75d2dc9, on a branch this sweep deleted).
• All fourteen findings dispositioned in the evidence store.

A cost I paid and would not repeat: round 1's fixes went in as an --amend, which rewrote the tree and destroyed the review fact, so verify-resolutions was refused and a full cumulative had to run. Round 2's fixes went in as a second commit and cost-of-commit priced them free — same fixes, no round. On a branch carrying a review fact, amend is the expensive option.

Note for the reviewer

prawduct-hook check-change-log-entry reports no-entry structurally on this repo: .prawduct/ is gitignored by owner ruling (.gitignore:41), so git cannot see .prawduct/change-log.md. Both entries were verified by hand — the private one at change-log.md:262, and the tracked public CHANGELOG.md entry under [Unreleased] / ### Internal. ### Internal is the correct tier: dev tooling and docs with no user-visible change.

[Jason-Vaughan]

Closing in favour of #214, which rebuilds the surviving half of this sweep on current main.

This branch is based on f68fe90 (#143) and main is now at 3547a65 — sixty PRs later. It conflicts in CHANGELOG.md, FEATURES.md and PROJECT-MAP.md, and rebasing would not have been the cheap part. Taken piece by piece:

• FEATURES.md — this branch removed ten TBD stubs. Main has since replaced them with 27 lines of real descriptions. Merging this now would delete that content, not restore it. Not carried.
• PROJECT-MAP.md — the defect is still on main, in a newer form: both the docs/ and tanglebrain/ bullets end mid-list with an unclosed parenthesis. But this branch's fix would not have held. It wrapped its descriptions across lines, and the project-map wrap-step reads the list line-by-line and drops continuation lines — that is exactly what severed the current entries (Session wrap on wrap/20260907001451-tanglebrain #155 wrote a correct three-line docs/ description, Session wrap on wrap/20260907020112-tanglebrain-builder #159's wrap-step deleted its two continuation lines). Repair the two PROJECT-MAP entries the wrap-step severed #214 repairs them as single physical lines and records why they must stay that way.
• .gitignore — the three coverage patterns are still missing on main. Carried into Repair the two PROJECT-MAP entries the wrap-step severed #214 unchanged.

The wrap-step bug itself is TangleClaw's and is filed separately.