refactor(perft): code-review polish for the perft rig + MCP server

[ber4444]

Applies six code-review findings from the merged perft/MCP work (#49). No behavior change to the green gate — all fixes are correctness/robustness/clarity polish. The shipped move generator is untouched; these fixes are to the rig and the MCP adapter, not the engine.

Findings addressed

1. Stale path in the divergence report — PerftVsStockfishTest wrote "Inspect the corresponding rule in app/src/commonMain/.../Move.kt", but the file moved to chess-core in the perft-tests migration. An agent reading build/perft-divergence.txt was pointed at a non-existent path (the exact "stale path after a module move" failure the docs hold up as the meta-lesson). Now points at chess-core/src/commonMain/.../Move.kt.
2. render() hid a filesystem write — split DivergenceReport.render() (now pure) from a new persist(rendered); call sites persist(...) then fail(...) explicitly instead of writing to disk as a side effect of building an assertNull message.
3. Timeout could masquerade as a divergence — on deadline/early-exit the rig (StockfishPerft) now throws instead of returning a partial divide, and the MCP adapter (StockfishDivider) reports a new DivideResult.timedOut and withholds the partial map. A slow engine can no longer look like a real mismatch.
4. Unbounded gate subprocess — run_perft_gate now bounds gradle (15 min default / 45 min deep) via a daemon stdout drainer + waitFor(timeout) + destroyForcibly(), so a wedged build can't hang the MCP tool call (and the agent) forever.
5. Duplicated depth cap — tool descriptions interpolate StockfishDivider.MAX_DEPTH / DEFAULT_TIMEOUT_MS instead of a separate MAX_DEPTH_CAP constant that could drift from the real clamp.
6. Unused import — removed kotlin.io.path.pathString.

Plus a docs commit: syncs the docs/perft.md MCP tool table with the new contract (gate timeout; stockfish_divide's timedOut), and adds a Future work note on using The Grand Chess Tree's ~28k-position corpus to broaden Oracle 1.

Verification (ran locally, this branch, Stockfish present)

• ./gradlew :perft-mcp:test --rerun-tasks — green after a fresh recompile. Covers the description assertions that require the literals "clamped to 6" / "120s" (now produced by interpolating the real constants) and formatDivideResult.
• ./gradlew :chess-core:desktopTest --tests "*Perft*" --rerun-tasks — green after a fresh recompile. Oracle 2 (PerftVsStockfishTest) actually ran (2 tests, 0 skipped, 0 failures) against /opt/homebrew/bin/stockfish; no perft-divergence.txt was written, consistent with a green gate.
• The new failure-only branches (rig-side timeout throw; render/persist split) are compile-verified and unit-covered, but not runtime-triggered by this green run — correctly, since nothing diverged or timed out.

🤖 Generated with Claude Code