fix: guard against episode storm stalling foreground sessions

[de1tydev]

Problem

Large merged episodes trigger a cascade of expensive post-processing (capture → reward → L2 induction → L3 abstraction → skill crystallization) that can stall OpenClaw and Hermes Agent foreground sessions. This is especially common in long development workflows where relation.classify consistently returns revision/follow_up, allowing a single episode to accumulate dozens or hundreds of turns.

Fixes #1755

Root Causes

1. No episode turn limit — episodes grow unbounded; the full L1→L2→L3→skill chain hits all at once when the topic finally ends
2. Synchronous classify in before_prompt_build — relation.classify() is an LLM call that blocks foreground prompt construction with no timeout
3. Unlimited background LLM concurrency — capture/reward/L2/L3/skill subscribers fire unlimited parallel LLM calls, starving the event loop

Changes

Fix 1: Episode turn hard limit (maxTurnsPerEpisode)

• New config: algorithm.session.maxTurnsPerEpisode (default 30, range 5–200)
• When an open episode reaches this turn count, the next turn forces a topic boundary regardless of relation classification
• Also applies when reopening recovered episodes

Fix 2: Relation classify timeout (classifyTimeoutMs)

• New config: algorithm.session.classifyTimeoutMs (default 5000ms, range 1000–30000)
• relation.classify() calls are wrapped with Promise.race against the timeout
• On timeout, defaults to new_task (safe conservative boundary)
• Prevents foreground prompt construction from blocking indefinitely

Fix 3: Background LLM concurrency semaphore (bgLlmConcurrency)

• New config: algorithm.session.bgLlmConcurrency (default 2, range 1–8)
• Shared semaphore gates all LLM calls from capture, reward, L2, L3, skill, and feedback subscribers
• Prevents event-loop starvation from concurrent background processing
• Capture's existing llmConcurrency (per-step α scoring) is unaffected — the semaphore only applies to the shared LLM client used by post-capture processing

New Files

• core/util/semaphore.ts — lightweight async semaphore
• core/util/rate-limited-llm.ts — transparent LLM client wrapper that acquires a semaphore permit per call

Files Modified

• core/config/schema.ts — 3 new config fields with JSDoc
• core/config/defaults.ts — defaults: maxTurns=30, classifyTimeout=5s, bgConcurrency=2
• core/pipeline/types.ts — SessionRoutingConfig extended
• core/pipeline/deps.ts — config extraction + semaphore wiring
• core/pipeline/orchestrator.ts — turn-limit guard + classify timeout wrapper

Testing

• tsc --noEmit passes (no type errors)
• All new config values have sensible defaults that preserve existing behavior for users who don't change them (the turn limit is the only behavior change: episodes that would have grown past 30 turns now get split)

[de1tydev]

Linked to #1755 — detailed root-cause analysis is in the issue comments.