Release LiRu-Bot v0.2.0 — clrsrc v1.1.1 (+~100 Elo) + connectivity hardening · clrsrc/liru-bot

v0.2.0 (2026-06-10)

clrsrc v1.1.1 (matefix + repfix + IIR-late): ~+100 Elo over v0.1.0's clrsrc v1.1.0
- matefix: fixes short-mate repetition shuffle (+48 Elo), closes game uEn2qBri class
- repfix: fixes long-mate 3-fold draw-detection in depth<=0 leaf, closes game lPG7cqDO class
- IIR-late: moves the Internal Iterative Reduction block after pruning gates — effective pruning
  depth is no longer reduced, yielding stronger pruning and +56 Elo

Three fixes that prevent mid-game connection losses from forfeiting games:

Fix 1a — Game-stream resubscribe (clock-aware): When the game stream drops (EOF or transport
error), the bot now attempts to reopen it with an exponential backoff (1s / 2s / 4s, hard cap 8 s,
total budget min(12 s, remaining_clock / 10)) rather than immediately abandoning the game.
On reconnect Lichess resends a gameFull event; the new gameFull reconnect arm replays the full
move list from that event so the board state is always canonical (idempotent — no double-move risk
even if our last move was already delivered before the drop).
Fix 2a — Stream-open Storm-Killer: get_event_stream / get_game_stream now use a single
HTTP attempt instead of the old with_backoff loop (100 ms retry every 5 s for up to 60 s on
failure). That old loop issued ~600 rapid requests on a transient blip, triggering Lichess's
/api/stream/event 429 cascade which killed all game streams account-wide. A single failing
attempt is now surfaced immediately; the reconnect cadence is owned by the outer loop.
Fix 2b — Event-stream 429 back-off: When the event-stream reconnect receives an HTTP 429
RateLimited, the bot now honors retry-after from the Lichess response (plus jitter) instead
of pressing ahead at the fixed 5 s exponential cadence. This prevents repeated re-triggering of
the 429 window.

Diversity brake: Matchmaking now enforces max_challenges_per_opponent_per_day (default: 5)
so a single online bot cannot monopolize the challenge queue. A soft weighting additionally
down-weights already-played opponents in favour of fresh bots when several are available.
Config key matchmaking.max_challenges_per_opponent_per_day (integer, 0 = unlimited).

Added movetime_cap_ms on the subprocess path: instead of forwarding raw clock times to the
engine via go wtime/btime, the bot now sends go movetime N where N = remaining/30 + inc.
This bounds the rare soft-inflation overshoot class (clrsrc's stability_factor could inflate
the soft limit on oscillating positions, causing single moves > 120 s → forfeit on slow hardware).

Per-move engine eval is now logged at DEBUG level (ply, bestmove, score, depth,
nodes, pv) when using the embedded engine backend. Useful for post-game analysis.