Parse human-draft picks and completion

## Problem

The parser only consumes bot-draft pick messages and enforces exactly one selected card. Premier, Traditional, and Pick-Two use `EventPlayerDraftMakePick`, and private Pick-Two evidence directly shows two `GrpIds` in one submission plus a distinct completion message.

## Proposed change

Parse human-draft pick submissions and completion according to the detected format/rules contract. Treat `GrpIds` as a collection, correlate submissions with pack notifications, and handle missing/duplicate/truncated records without format guessing from cardinality when metadata is present.

## Acceptance Criteria

- [ ] Premier and Traditional one-card submissions produce one-card logical picks.
- [ ] Pick-Two two-card submissions retain both IDs in one logical pick.
- [ ] Explicit event metadata wins over payload-cardinality heuristics.
- [ ] Human-draft completion is recognized from observed messages without applying Quick Draft's 42-card inference globally.
- [ ] Missing submissions, failed-pick gaps, duplicate responses, and truncated envelopes do not crash or fabricate picks.
- [ ] Unknown human-draft events fail safely without contaminating the active draft.
- [ ] Hand-authored synthetic fixtures cover all supported message shapes; raw logs remain private.
- [ ] Existing Quick Draft behavior and focused parser tests pass, with a changelog entry.

## Scope notes

No persistence redesign beyond Child 2 and no frontend behavior.

Start with the parser and existing focused parser tests, then use hand-authored synthetic fixtures to cover the supported message shapes described in the acceptance criteria. Done means human-draft submissions and completion are handled safely, missing or malformed records do not create false picks, existing Quick Draft behavior still passes, and the changelog is updated.

索引モデルが issue の本文から書いたものです。