One page for the unsupervised-learning goal, instead of a branch

[syncytium2]

Work toward a label-free detector had scattered across handoffs, proposals, review records, todos, a draft branch and darkroom run folders, and the goal itself was getting lost. Tony asked whether a long-lived `unsupervised_learning` branch would keep sessions on target.

Why not a branch: sessions start from `main`, the briefing and `docs/INDEX.md` read `main`, and a branch collects commits rather than saying what is settled. A branch would have added one more place to look.

What this adds instead

• `docs/goals/README.md` sets the convention: one page per goal, `Goal:` on the board claim, a branch-name prefix, and the page updated in the same PR as each result.
• `docs/goals/unsupervised-learning.md` is the first goal page. It covers where the goal stands, settled findings with their strength and source, dropped approaches, the decisions waiting on Tony, open work a session can do without a ruling, and where the work lives.
• There are pointers from `CLAUDE.md`, `docs/INDEX.md` (a new lead row) and `docs/model_track.md`.

Found while reconciling

• The screen's machinery, its reviewed report record and its latest handoff exist only on draft The surrogate screen's machinery: twelve candidates, six controls, three yardsticks #530.
• `surrogate-screen-plan`, `surrogate-field-ruled` and `the-conditioned-run` show as unpushed in the briefing, but they were squash-merged as Measure tonight, decide tomorrow: the surrogate-screen plan, its two-round review, and the darkroom claim #529, The whole field enters the screen: Tony ruled it, and Stella's figure disagrees with Stella's conclusion #526 and The conditioned run, and the eight questions it cannot answer for itself #514. Nothing is lost.

This is working material, not murderboarded, with the same standing as `docs/pipeline.md`. Docs only. The INDEX, sapper and quote checks pass, as do the 49 test files that read docs.

🤖 Generated with Claude Code