feat: add pi-coding-agent integration

[bytes032]

Summary

This PR finalizes PI integration on top of Entire’s generic lifecycle architecture while preserving PI-specific tree transcript behavior.

PI now follows the same integration flow as Claude/Gemini:

ParseHookEvent() -> Event -> DispatchLifecycleEvent()

PI-specific differences (tree branches + extension hooks) remain intentional.

What Changed

1. PI lifecycle mapping on new architecture

• session-start -> SessionStart
• user-prompt-submit -> TurnStart
• stop -> TurnEnd
• before-compact -> Compaction
• session-end -> SessionEnd
• leaf_id from PI stop is captured and persisted for active-branch correctness.

2. PI tree-aware transcript behavior preserved

• Active-branch (leaf-aware) extraction is retained for:
• prompts
• modified files
• token usage
• summary extraction
• Leaf hints are persisted so post-hoc reads resolve the correct branch.

3. PI hook robustness improvements

• Added non-blocking timeout protection for hook subprocess execution.
• Improved session_switch / session_fork previous-session end handling by deriving prior session ID from previousSessionFile when needed.

4. Hook registration parity

• PI implements HookHandler (GetHookNames) so entire hooks pi <verb> subcommands are registered consistently like other agents.

5. PI-aware condensation correctness

• Condensation paths now use PI leaf context so prompt/token/summary/context extraction is scoped to the active branch.
• This avoids reading abandoned branches from append-only tree transcripts.

6. PI-only stop flush mitigation

• On stop, PI now waits briefly (best-effort, bounded) for transcript flush/non-empty content before emitting TurnEnd.
• This is PI-only and does not alter shared core lifecycle behavior.

Why these changes are needed

Without these fixes, PI metadata can drift in branched sessions:

• prompts/context from abandoned branches
• token overcount across non-active branches
• summary mismatch vs active branch
• checkpoint metadata not matching user-visible branch state

Known Limitation + Strict Core Fix Path

Current limitation (shared core behavior):

• turn-end full-transcript finalize remains best-effort
• if finalize fails at turn-end, a checkpoint may retain provisional transcript content

How to make this strict (core change, not in this PR):

1. Do not clear pending TurnCheckpointIDs when finalize fails.
2. Retry UpdateCommitted on subsequent safe lifecycle points (TurnEnd / SessionEnd).
3. Clear pending IDs only after successful full-transcript update.

This would provide eventual full-transcript consistency as a hard guarantee.

PI-Specific Differences (Intentional)

• PI uses extension-based hook integration (.pi/extensions/entire/index.ts), not JSON settings-based hook registration like Claude/Gemini.
• This is intentional and matches PI’s native extension model.

Validation

• go test ./cmd/entire/cli/agent/pi ./cmd/entire/cli/summarize ./cmd/entire/cli/strategy
• go test ./cmd/entire/cli/...
• All passing.

[bytes032]

@Soph @gtrrz-victor @khaong this should be ready for review

[khaong]

Thanks for putting this together! We're currently reworking the agent interface (#281) to make integrations simpler and more consistent, so we're holding agent PRs until that lands — otherwise a lot of this would need to be redone against the new interface.

We'll update #281 when it's ready for contributions.