docs: correct the retired develop flow, right-size the run rules, add RN test traps

[taylorcox75]

Housekeeping on AGENTS.md / docs/RELEASING.md after noticing during #245 that the doc's branching instructions no longer match the repo.

The doc was actively misleading

develop doesn't exist. AGENTS.md mandated it in ten places — "Never commit to develop", "Always cut from develop — there is no hotfix exception", gh pr create --base develop. There's no such branch locally or on origin, and no PR has ever targeted it. Everything since #219 goes straight to main; #218 and earlier used preview. An agent following the doc gets a hard failure on git switch develop, or opens a PR against a nonexistent base.

Rewritten around main, with an explicit "develop and preview are retired" note so the next reader doesn't have to re-derive it, and a row for the release/vX.Y.ZZ branches that are used occasionally.

docs/RELEASING.md's changelog decision tree was half-dead for the same reason — the "On develop → .TESTFLIGHT placeholder" branch can never fire (grep -c TESTFLIGHT constants/changelog.ts → 0). Removed. The main rules it sat next to are unchanged.

The run-cost model was off by ~25x

	claimed	measured
one util suite	~18s	1.2s
tsc warm / cold	~25s / ~85s	0.8s / 3.2s
full npm test (1095 tests)	—	4.9s
npm run lint	—	6.4s

The doc built a three-tier ceremony ("exactly three moments to run anything, and no others") around what is a ~12 second batch — so it was discouraging verification to save almost nothing. It had a real cost: in #245 I followed "mid-task → run nothing" and shipped tests that failed in ways a one-second run would have caught immediately.

The constraint that actually binds is context, not wall time — npm test prints hundreds of lines of unrelated act() warnings and Animated stack traces from Confetti/PathAutocompleteInput, and all of it persists in the conversation. So the rule is now "run the narrow check freely, but pipe it through tail." The commit-time batch is unchanged.

Two rn test traps documented

Both cost roughly an hour to rediscover in #245:

• A synchronous act(() => …) around a promise-returning call silently swallows the state update — the component never re-renders, no warning, and it looks exactly like a product bug. Call it bare and let waitFor observe.
• render() returns a promise, so an outside-provider throw needs await expect(render(…)).rejects.toThrow(…).

Smaller corrections

• Critical Rule 6's Alert deviations list omitted ConfirmModal itself — it's nine calls across eight files, not eight.
• File Index gains SkeletonTorrentDetail; ServerContext entry documents isReconnecting, and TorrentContext documents how the recovery flag is cleared (both from Fix reconnect UI after a long background (skeleton instead of stale data + auth toast) #245).

Verification

Docs only — no code touched, so no test/lint impact. Claims were checked against the repo rather than assumed: branch list, gh pr list base branches, workflow triggers, grep -c TESTFLIGHT, the Alert call sites, and the timings above were each measured. I also verified the File Index is 100% complete — every file in utils/, hooks/, context/, services/, services/api/, components/, modules/, constants/ is documented; that claim is holding up well and is left alone.

🤖 Generated with Claude Code