feat(openclaw): drop redundant before_agent_start + add debugPerfTiming

[nicoloboschi]

Summary

Two small, unrelated openclaw plugin changes bundled in one PR.

#1354 — drop before_agent_start registration

The hook's body did nothing observable: a resolveAndCacheIdentity() call and a debug log. The same identity resolution already happens in:

• before_dispatch — earlier in the inbound channel path, populates the cache.
• before_prompt_build — re-resolves before recall, can also infer senderId from the raw prompt when ctx lacks it.
• agent_end — re-resolves before retain.

Subscribing was duplicate work on the hot path. Removed the registration and the \"agent_start\" member of the logSkipOnce operation union.

Also fixed a stale comment in `stripMemoryTags` doc that referenced `before_agent_start` — the actual injection happens in `before_prompt_build`.

#1406 — opt-in `debugPerfTiming` flag

Adds `debugPerfTiming?: boolean` plugin config (default `false`). When on, the plugin emits one info-level perf line per recall path and per retain path:

```
perf: before_prompt_build hook_total=4200ms recall_main=3800ms source=fresh results=3
perf: agent_end hook_total=1200ms retain=1100ms outcome=ok bank=main messages=4
```

This lets users diagnose latency without patching the compiled dist (which is what the issue reporter resorted to). Specifics:

• Info-level, not debug-level, despite the flag name. Perf data is small-volume / high-signal — gating it behind `debug: true` would silently produce no output for users who only want timing. Documented.
• Only emitted on the recall/retain RPC paths, not on the many fast skip-filter early returns. Those are sub-millisecond and not interesting.
• `source=fresh|reused` reflects in-flight recall dedup at the hash key.
• `outcome=ok|queued|error` reflects whether retain succeeded inline, was queued for retry, or failed outright (queue-less mode).
• The pure formatter `formatHookPerf()` is exported and unit-tested for the rendering contract.

Test plan

• `npm test` — 212 tests pass (5 new for `formatHookPerf` + `debugPerfTiming` parsing).
• `npm run build` — tsc clean.
• `./scripts/hooks/lint.sh` — clean.
• Manually enable `debugPerfTiming: true` in a real openclaw setup, verify perf lines appear with sensible values for both hooks.

Out of scope

• feat(openclaw): Make agent_end retain non-blocking to reduce hook latency #1407 (non-blocking retain) closed separately as already-solved by the existing server async contract — the plugin sends `async: true`, server enqueues and returns immediately, full processing is in a background worker.
• feat(openclaw): Option to skip retention of low-value transcripts (chitchat, noise, pure tool calls) #1405 (retain quality gate) deferred — see issue triage discussion for the counter-proposal.

Closes #1354
Closes #1406