feat(puzzle): add custom puzzle builder validation and tamper-evident replay verification

[nafiuishaaq]

Summary

Adds custom puzzle authoring and tamper-evident replay verification to the
puzzle engine service.

What this adds

src/puzzle/puzzle.service.ts:

• Custom puzzle builder (Custom Puzzle Builder and Submission System #435)
  • validatePuzzleDefinition() rejects malformed boards with field-level
    errors ({ field, code, message }) rather than throwing, so the builder UI
    can attach each message to the input that caused it. It checks grid bounds,
    that cells and solution both match width * height, that every value is
    in allowedValues, and that neither board nor solution is trivial.
  • estimateDifficulty() returns a score, a band (easy/medium/hard/
    expert) and the raw inputs it used, so a preview can be shown and
    recalibrated later. It is a transparent heuristic — log-scaled board area,
    distinct-value variety, and solution run density — not a solver.
  • submitCustomPuzzle() runs validation and returns a moderation envelope. A
    valid puzzle is never published directly: it always enters pending_review
    with a definitionDigest over the canonicalised definition, so a later
    moderation decision can be checked against the exact bytes that were
    validated.
  • moderateSubmission() moves a submission to approved/rejected, refusing
    to moderate a non-pending submission or one reviewed by its own author.
• Replay recording and verification (Puzzle Replay Recording and Sharing #436)
  • recordReplayAction() requires contiguous sequence numbers and
    non-decreasing timestamps, so a replay cannot skip or reorder moves, and
    folds the previous chain hash into each action digest.
  • finalizeReplay() seals a replay into a summary (action count, duration,
    hints used, chain hash, submitted-value hash) for playback and sharing.
  • verifyReplay() checks three things before a result is trusted: the actions
    still hash to the sealed chain digest (CHAIN_TAMPERED), the submitted
    value matches the recorded solution hash (SOLUTION_MISMATCH), and the run
    was not faster than MIN_PLAUSIBLE_SOLVE_MS (IMPLAUSIBLE_SOLVE_TIME).
• createPuzzle() now uses a shared private hash() helper instead of
  inlining createHash, which the new code also needs.

Notes

• No new dependency was injected. Adding a TypeORM repository would require
  editing puzzle.module.ts, a second file, so the builder and replay work is
  built on the already-injected SorobanService and node:crypto. As a result
  submissions and replays are returned as envelopes rather than persisted;
  wiring them to storage is the follow-up this change deliberately leaves out.
• The replay chain is tamper-evident, not tamper-proof. Recomputing the
  chain detects edits to a stored replay, but an attacker who can rewrite the
  whole record can also rewrite the chain. Sealing the final digest somewhere
  append-only (or signing it) is what would actually prevent forgery.
• estimateDifficulty was checked against small/medium/large boards while
  developing and now yields easy/medium/hard respectively. An earlier version
  normalised run density by area / 2, which let a 2x2 board max out that term
  and score "medium"; it is normalised by area now.
• The difficulty weights are unvalidated against real solve data, so a preview
  will disagree with players until it is calibrated. inputs is returned
  precisely so that calibration is possible later.
• Replay compression, sharing, comments, analytics and expiry are not
  included — they live in src/replay/ and src/puzzle/session-replay.service.ts,
  which are separate files.
• No unit tests are added here, per the current scope of this change.

Issues

Closes #435
Closes #436
Closes #437
Closes #438

[drips-wave]

@nafiuishaaq Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits