Ship track_locality.py: reproducible blacklist + whitelist track-locality measurement (#11)

[sepehrsadri]

Progress on #11.

What this does

1. Adds experiments/track_locality.py, implementing both approaches the issue asks
   for, reported side by side:
   • Blacklist — hash each track's whole XML subtree, dropping a frozen, named,
     commented exclusion set (TRACK_BLACKLIST_ELEMENTS / TRACK_BLACKLIST_ATTRS). Every
     entry has a one-line reason and, where relevant, a pointer to which observed diff
     motivated it.
   • Whitelist (the "bonus, and better" ask, and AGENTS.md's own stated preference) —
     extract only the fields that define "the track" (name, volume, pan, on/off, output
     routing, device chain, clip sample/position/mute) and hash that instead.
   • --chain 'Backup/*.als' support, reusing als_semantic_diff.py's exact
     autosave-lineage-grouping convention (the AUTOSAVE_NAME regex and split-by-stem
     logic) rather than inventing a second one.
2. Runs it against a real local Ableton chain and reports the actual output — see
   "Measurement" below.
3. Updates docs/EXPERIMENTS.md §3 with that output, removes the "provisional" warning,
   and is explicit that this is a new measurement on smaller, different material than
   whatever produced the retracted "median 3" figure.
4. Adds tests/test_track_locality.py, matching this repo's convention of a synthetic-
   fixture test file per experiments/ script. It pins: identity (no false positives),
   real content changes caught by both methods, view-state/warp-marker noise caught by
   neither, and — the one I actually expected might fail before writing it — that editing
   one track does not falsely flag an untouched track just because the edit shifted a
   positional id counter (AutomationTarget/ModulationTarget) elsewhere in the file.
   All 7 passed on first run, which is the exclusion set doing its job on synthetic
   material as well as real.

Measurement

Sample: 10 consecutive autosaves (9 pairs) of a 28-track Ableton Live set, one local
chain, this environment.

Method	Median	Range	Per-pair counts (of 28)
Blacklist (frozen exclusion set)	9	0–15	5, 9, 4, 0, 15, 1, 11, 11, 12
Whitelist (named fields)	2	0–4	4, 3, 4, 0, 4, 0, 2, 0, 0

The two disagree by >4x on this material too, which is the whole point of shipping both:
on two of the nine pairs, als_semantic_diff.py independently reports "no musical change
detected", the whitelist agrees (0 of 28 changed on both), and the blacklist reports 11
and 12 changed tracks on those same two pairs — pure residual churn (chiefly warp-marker
re-analysis and FileRef/AuPreset/AutomationTarget id renumbering the exclusion set
doesn't fully catch). That is the exact "blacklists leak" warning in AGENTS.md,
reproduced on new material rather than just asserted.

The qualitative claim survives either way — even the less flattering, blacklist-inflated
median (9 of 28) still means two-thirds of a typical save is untouched.

Reproduce:

python3 experiments/track_locality.py --chain 'path/to/YourProject/Backup/*.als'

What this PR deliberately does NOT do — and why I'm not asking to close #11

The issue's number (median 3 of 28, provisional) and the "independent re-implementation"
table in the issue body both trace back to the original 26 GB / 30-project corpus,
which lived on a different machine and is not available in this environment. I have a
different, smaller, local chain (28 tracks, 9 pairs, one lineage) — genuinely useful for
validating the method, but not the same material, and not large enough to stand in as
"the" corrected headline number for a document that explicitly wants figures backed by
real, full-scale material (see the Material table at the top of EXPERIMENTS.md).

So this PR:

• retracts the old "median 3" figure as unreproducible (not silently — the doc says so
  and says why),
• publishes an honestly-scoped new number instead of a bigger, unverifiable claim,
• ships the script + frozen exclusion set the issue actually asked for, so whoever does
  have access to the original corpus (or a comparably-sized one) can now get a real
  answer with one command instead of reverse-engineering an exclusion set from prose.

Recommendation: keep #11 open, re-scoped (or split into a follow-up) to "run
track_locality.py against the original/a comparably-sized corpus and publish that
number." I don't have that material here, and inventing a number to close the issue
would repeat exactly the mistake #11 was filed over.

Verification

• ~/.cargo/bin/uvx ruff check experiments/ tests/ — clean
• python3 .github/workflows/scripts/check_personal_paths.py — clean
• python3 -m pytest tests/ -q — 293 passed, 13 loudly-skipped (opt-in real-fixture
  tests, expected in this environment)

🤖 Generated with Claude Code