Root cause: currentDrillGame state changes on every player/Maia move
(updating moves array and playerMoveCount). The cancellation effect
depended on currentDrillGame, so every move triggered cleanup which
set backgroundCancelledRef=true, killing the loop immediately.

Fix: track drill identity by ID instead of the full state object.
Only cancel when a genuinely new drill starts. Also stop background
analysis when post-drill analysis begins to avoid stockfish contention.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Background loop runs both Maia + Stockfish on drill positions as
they're played. When the drill ends, ensureDrillAnalysis:
1. Sets cancelled flag (loop exits after current position)
2. Calls stockfish.stopEvaluation() to abort any in-progress eval
3. Awaits the loop promise to ensure full exit
4. Only then starts its own stockfish work

This avoids the concurrent generator bug where the old generator's
finally block would clobber the new generator's engine.listen
callback, causing the post-drill analysis to hang.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

React runs useEffects in declaration order. The separate cancellation
effect was firing AFTER the enqueue effect, clearing the queue that
was just populated. Merged both into a single effect so drill-change
reset and node enqueuing happen atomically.

Also handle drill-cleared case (cancel background when no active drill).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Detailed [bg] logs at each stage to diagnose why background analysis
doesn't run during drill play.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The queue/loop pattern had too many failure modes (running flags,
queue clearing, effect ordering). Replaced with a simple promise
chain: each new drill node gets its analysis chained onto a single
ref promise. No queue, no running flag, no separate loop callback.

- bgChainRef: single promise chain, analysis tasks appended via .then()
- bgAnalyzedFensRef: Set tracks which FENs have been scheduled
- Drill change: cancel flag + chain reset
- stopBackgroundAnalysis: sets cancel + stops stockfish + awaits chain

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Two fixes:
1. Wrap each chain step in try/catch so a single error doesn't reject
   the entire promise chain
2. Don't await the chain when stopping — just signal cancellation and
   stop stockfish, then proceed. Awaiting was hanging when a step was
   stuck (e.g., Maia ONNX inference or stockfish generator)

Also added more detailed [bg] console logs at each stage.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Root cause: useAnalysisController's useEngineAnalysis hook auto-runs
stockfish on every currentNode change. During drill play, each move
synced the node to the analysis controller, which called
stockfish.streamEvaluations() — killing the background analysis's
in-progress evaluation (the new call's stopEvaluation() terminates
the old generator).

Fix: only sync currentNode to the analysis controller when analysis
is actually active (post-drill continue-analyzing mode), not during
active drill play.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The analysis controller's useEngineAnalysis hook has its own internal
tree controller that navigates the drill's game tree. Even though we
stopped syncing currentNode, the internal controller still tracked
tree changes and fired stockfish.streamEvaluations() — which calls
stopEvaluation() and kills the background analysis's in-progress
evaluation (explaining depth 8 instead of 18).

Added an 'enabled' parameter to useEngineAnalysis and
useAnalysisController. During drill play (analysisEnabled=false,
continueAnalyzingMode=false), both Maia and Stockfish effects in
useEngineAnalysis are completely skipped.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>