---
description: Bundle, bracket, and live-data invariants before opening a PR
globs: packages/core/**/*
alwaysApply: false
---

# Bundle & bracket PR readiness

Before marking a data-heavy feature merge-ready, run **`pnpm -r build && pnpm -r typecheck && pnpm -r test && pnpm lint`** (mirror CI — build first because cli/mcp import `@claudinho/core` from dist; `build` alone does not typecheck).

## Bundled schedule invariant

`schedule.*.json` is a **resultless skeleton**: no scores/status, no confirmed nations in knockout slots.

- `sanitizeBundledFixture` must restore topology placeholders for KO home/away.
- `gen:schedule` must **fail** if any knockout fixture carries a real nation flag, and at once,
  writing nothing, on a window that failed, carried no account, was not whole, or stated another
  season (or none), and on a fixture served by two windows; its expected counts are named facts (`GROUPS`, `TEAMS_PER_GROUP`) and core's
  `EXPECTED_KNOCKOUT_COUNTS`, never literals or this run's own topology.
- Test degraded `getBracket()` on the bundle: seed/winner slots stay `tbd`, never `confirmed`.

## Knockout resolution

- Advance winners from **score OR `winnerCode`** (ESPN `competitor.winner` for penalties).
- Always test FT draw + `winnerCode` — routine in knockouts.
- Group slots project from live standings once **≥1 match played** (`hasGroupStarted`); **`confirmed`** when every team has played a full round-robin (`played >= n - 1` for `n` teams); **`(proj.)`** only mid-group. TBD at 0 games or when standings are degraded.

## Topology indexing

- Assign each bundled match's 1-based index within its stage by ascending numeric ESPN event id
  (`orderByBracketIndex` in `packages/core/src/bracket/build.ts`), never by kickoff sort order.
- **The live fixture's pairing wins.** ESPN's bundled placeholder winner/loser refs do not
  reliably describe its actual feeder assignment. Use the participant served for that fixture;
  project from a bundled feeder ref only when the target fixture is absent from the merged set.
- Keep the disagreeing-feeder `P1 GUARD` in `packages/core/test/bracket-resolve.test.ts` green,
  and cover the feeder fallback when the target fixture is absent. See `AGENTS.md` → Conventions.

## Live hybrid paths

- When knockout fetch fails but standings succeed: set `source` from standings provider.
- Share/MCP: inner `formatBracketList({ footer: false })`, outer layer adds attribution/notices **once**.
- CLI/MCP wrappers: option types must match the formatter (`ShareBracketOptions`, not `ShareSnippetOptions`).
- **Surface parity:** see `.cursor/rules/surface-parity.mdc` — every bracket formatter path passes `tz` + `locale`; statusline knockouts indirect-resolve via cached `getKnockoutFixtures` (hot path fail-closed).

## Pre-merge QA

After `pnpm -r build`, run **`pnpm release:qa`** (`scripts/release-qa.sh`). Eyeball every section; tripwires at the end encode the 0.8.x bracket regressions (calendar month, tz threading, share disclaimer).

## Adversarial tests (add when touching bracket/schedule)

| Scenario | Assert |
|----------|--------|
| FT level + `winnerCode` | Winner advances |
| `fetchWindow` throws, standings OK | Projected groups + `source` set |
| Bundle-only bracket | No real flags in R32+ slots |
| Group at 0 games played | Slots stay `tbd`, no `(proj.)` |
| Group at 2/3 played, standings live | Leaders show flags + `(proj.)` |
| Group fully played (round-robin complete) | Winner + runner-up `confirmed`, no `(proj.)` |
| `toolGetBracket` `tz: UTC` vs `America/Mexico_City` | Calendar date/time differs on cross-midnight kickoff |
| Knockout kickoffs span >1 week | Output includes month + day (not weekday-only) |
| `formatShareBracket` social + compact | Same date format as CLI list |
| `formatShareBracket` / `toolGetBracket` | Each disclaimer line once |
| Statusline knockout countdown (cached fixtures) | Real nations or `⚽ —`; never `🏳️` |
| Statusline stale cache during live KO | `live · syncing…` without `🏳️`; resolved pairing when cache has it |
| Empty fixtures cache at phase boundary | Re-polls after short TTL (~60s), not 15min |