docs: require a repro bundle with every fix

[FelipeDefensor]

Why

Human verification of AI-written fixes is our delivery bottleneck, and most of that time is scenario setup — clone the PR, launch TiLiA, create a timeline, create components — before any judging can start.

This adds a ## Delivering a fix for human testing section to CLAUDE.md requiring every fix to ship a repro bundle: a .tla fixture, a copy-paste launch command, and three lines of acceptance criteria.

No code change needed

The mechanics already exist:

• tilia /abs/path/to/file.tla works today — positional argument in boot.setup_parser.
• A fixture with "media_path": "" plus a "media length" in media_metadata opens with no prompt and no error (App._setup_file_media returns early on an empty path), so fixtures need no bundled media.
• The CLI already has script <path> + save <path>.tla for generating fixtures reproducibly.

Two guardrails

These matter more than the convenience:

• The fixture must reproduce the bug on the base branch. Fix and fixture are usually written by the same author, so a fixture that only ever shows the fixed state can encode the same wrong mental model as the fix and prove nothing. Stashing the fix and relaunching the same file must show the broken behavior.
• The .tla is always deleted before merge. The file format changes and committed fixtures go stale silently. The durable artifact is the generator (CLI script or pytest), promoted into the suite as a regression test when the scenario is worth keeping.

Making the commands actually runnable

Building a bundle for #450 turned up three ways the original guidance produced commands that don't run. All three are now spelled out in the section:

• Paths handed to TiLiA must be absolute. Outside ENVIRONMENT=prod, dirs.setup_dirs chdirs into the tilia package, and boot() parses its arguments before that chdir. So tilia repro/x.tla passes argparse validation and then fails with "File not found", and a CLI save repro/x.tla writes inside tilia/ without saying so.
• The launch command has to go through the project environment. A bare tilia resolves to nothing unless .venv is active. The recommended form is uv run tilia "$PWD/repro/<issue>.tla" from the repo root — absolute after expansion, and portable without hard-coding anyone's checkout.
• gh pr checkout <N> does not belong in the copy-paste line. It aborts when the head branch is checked out in another worktree, its fast-forward fails once the local branch has diverged, and --force resets the branch and discards the reviewer's local edits — which they may have made deliberately (e.g. to reach a code path the fixture can't trigger). It is one-time setup plus a warning, not part of the repeatable command.

Known limitation

CLI coverage is uneven, so the CLI-script route doesn't fit every kind:

• timelines add covers hierarchy, marker, beat, score, range and audiowave — not harmony, pdf or slider.
• components registers only the beat subparser, so it is not a general way to populate a timeline.
• timelines import is usually the better route when the fixture needs components: marker and hierarchy (CSV, by-time or by-measure), beat (CSV), score (MusicXML).

Anything outside those lists falls back to a throwaway pytest that builds state via commands.execute(...) and saves via commands.execute("file.save", ...). A separate PR widening CLI coverage is in progress.

Bugs needing real media (playback, audiowave, video, PDF) can't ship a portable fixture at all — the guideline says to ask for a file rather than ship one that can't reproduce the problem. YouTube-backed bugs are the exception: the URL lives in media_path, so the fixture stays portable even though it isn't media-free.