Fix #1897 recovery replay request storm

[TianchunGu]

Summary

Follow-up for #1897, stacked on top of #1938.

#1899 / #1938 stop the terminal-provider-error retry storm at the LLM facade. This PR adds a second guardrail for the capture pipeline: startup recovery of dirty closed episodes can replay a large historical episode, misclassify recovered steps as orphan traces, and then trigger many reflect/summarize LLM calls.

This patch:

• fetches all episode traces for capture reconciliation instead of using the paginated viewer list() path;
• matches reflect replay steps by stable content signature, with a timing-insensitive fallback for recovered tool-call traces that lack startedAt/endedAt;
• skips recovered replay orphan inserts by default (maxRecoveryOrphanInserts: 0) to avoid duplicating historical traces;
• avoids LLM summarization for any recovered replay orphan insert that an operator explicitly allows;
• adds a hard per-episode reflect LLM budget (maxReflectLlmCalls, default 128);
• stops remaining reflect LLM attempts after terminal/circuit-open provider errors.

The goal is to keep normal short-episode reflection behavior intact while preventing a single dirty recovered episode from producing unbounded paid LLM calls.

Validation

From apps/memos-local-plugin:

npx vitest run tests/unit/capture/capture.test.ts tests/unit/capture/capture-batch.test.ts tests/unit/capture/normalizer.test.ts tests/unit/llm/client.test.ts
# Test Files 4 passed (4)
# Tests 57 passed (57)

npm run lint
# tsc -p tsconfig.json --noEmit
# passed

New regression coverage:

• recovered replay matches tool traces by payload when timestamps drift;
• startup-recovered replay orphans are not inserted by default;
• reflect-phase LLM calls are capped per episode.

Related: #1897, #1899, #1938