docs: truth up architecture, write-gating, and conventions

[amrtgaber]

Summary

Close-out documentation sweep after the generalization work — every drifted passage found by auditing the docs against current code:

Pre-existing drift (predates this week):

• README's architecture section still described the LISTEN/NOTIFY nudge spine that Introduce PgBouncer connection pooling (DB conn ceiling) #196 replaced: the worker "emits a pg_notify" and the api "runs a dedicated LISTEN connection". Rewritten to the real design — the worker bumps nudge_versions in-transaction (now tagged with the touched tournaments), each api instance polls it through the pooled engine, no session-pinned connection anywhere. Diagram labels updated; the api bullet's stale min=1 max=10 replaced with the event_mode-driven reality.
• README's Authentication section said owner rows are "inserted directly (SQL) for now; an API … is planned" — the owners API has existed for a while and is listed ten lines above. Replaced with the truth, plus a sentence on the one write that is still out-of-band by design: tournament_creators grants.
• LISTENER_ENABLED descriptions (README env table + app/config.py comment) updated to "nudge poll loop + SSE fan-out"; a note records that the env var name predates Introduce PgBouncer connection pooling (DB conn ceiling) #196 and stays as deployment contract.

Drift from this week's changes:

• README write-endpoints: POST /v1/tournaments no longer says "any authenticated user may" — it documents the creator allowlist and the can_create_tournaments gate; the PATCH bullet lists the real editable field set (rank_by included).
• CLAUDE.md: NudgeVersion + TournamentCreator added to the entity groups, and two new convention bullets — Tournament write guards (creator allowlist, the SUPPORTED_TOURNAMENT_LEADERBOARD_IDS id-allowlist-not-name-check lesson, the two caps) and SSE nudge payload (tournament_ids contract and per-tick scoping semantics).

Docs-only plus one comment; suite still 600 passed.

Deliberately NOT touched: infra/README.md's cost-baseline paragraph still describes the retired db-f1-micro era — cost numbers should only be rewritten against billing actuals, so that's flagged for a future cost-audit pass rather than hand-edited here.