fix(acp): implement session/delete so Zed Trash removes whale sessions

[reneleonhardt]

Summary

Zed gates thread deletion on agentCapabilities.sessionCapabilities.delete + session/delete; whale-acp advertised neither → agent-panel Trash was a cosmetic no-op (zed-industries/zed#51912).

Changes

• Capability: initialize advertises sessionCapabilities {list:{}, delete:{}}.
• session/delete: params required; sessionId validated (isSafeSessionID) before any file access. Live session: refuses in-flight prompts (every .jsonl writer is turn-scoped → file resurrection unreachable), else drops from map + closes runtime off-path. Unknown id → idempotent success.
• File removal: primary <id>.jsonl fatal on failure; sidecars best-effort (meta.json, .approval_events.jsonl, .tool_input_events.jsonl, .approvals.json, stale .jsonl.tmp).
• Hardening: prompt registration re-resolves the session under lock — closes the lookup→register gap where delete/eviction could leave a prompt on a closed runtime, resurrecting the .jsonl.

Validation

• go build ./..., go vet, gofmt clean; full suite green except internal/tools (pre-existing TestRunShellBackgroundDoesNotPanic 10m timeout).
• -race on internal/acp + internal/session + internal/store.
• Tests: advertise; happy path (artifacts gone, list excludes); unknown id; 9 invalid ids/params; live idle (Close fires); in-flight refusal; post-delete load replays nothing; fatal .jsonl failure; dir-as-file; _meta ignored; -race delete-vs-list concurrency.

Review guidance

• handleSessionDelete + removeSessionFiles (handler.go): in-flight guard, off-path close, fatal-vs-best-effort split.
• adapter.go registration re-check: closes the delete/eviction gap.

User-visible impact

• Trash on archived whale threads now deletes them.
• On success, Zed re-lists its session list itself (SessionListUpdate::Refresh after delete_session succeeds), so the thread disappears from the archive.
• Agent pushes no list-change notification — ACP has none; the client owns list refresh.
• Server-side delete is durable + idempotent, so a non-refreshing client only shows a stale thread until its next session/list.

Breaking changes

None. New method session/delete; capability addition only affects clients that check it.

Reproduction

1. initialize → sessionCapabilities.delete advertised.
2. session/new + session/prompt; archive; session/delete → {}; session/list → session gone, artifacts removed on disk.

Developed with carefully directed, manually reviewed AI assistance.

[shayne-snap]

Blocking: This hard-delete path leaves several session-scoped sidecars on disk: .state.json, .todo.json, .user_input.json, and .goal.json (see the corresponding state helpers under internal/session). Those files can retain the session mode, todos, pending user-input questions, and goal/objective after Zed reports the thread deleted.

Please include them in this cleanup list and extend the deletion test to create and assert removal of each.

[reneleonhardt]

Thanks for the careful review — confirmed. session/delete (removeSessionFiles) removed the primary .jsonl plus the meta/telemetry/approvals sidecars, but left the session-state files persisted by the helpers under internal/session. I cross-checked every session-scoped path in the repo; the full missing set is exactly the four you listed:

• .state.json (mode) — internal/session/mode_state.go
• .todo.json (todos) — internal/session/todo_state.go
• .user_input.json (pending user-input questions) — internal/session/user_input_state.go
• .goal.json (goal/objective) — internal/session/goal_state.go

(internal/app/commands_doctor.go independently lists meta/state/todo/user_input; .goal.json is confirmed by goal_state.go.)

The fix adds all four to removeSessionFiles using the same SanitizeSessionID(id)+suffix path construction as the helpers. TestSessionDeleteRemovesPersistedSession now creates each of the four sidecars and asserts they are gone after delete, and assertSessionArtifactsGone covers them as well. The full internal/acp suite passes.

For completeness, the sweep also verified the remaining per-session artifacts are out of scope for this path: promoted_tools.json is written only by the main app into a per-session subdirectory (sessionsDir/<id>/promoted_tools.json), and tool-result payloads live under ~/.whale/tool-results/<id>/ — the ACP handler creates neither; it keeps flat files in its sessions dir.

[reneleonhardt]

Commits are only separated for reviewability, you can Squash and merge.
https://docs.github.com/en/pull-requests/reference/pull-request-merges#squash-and-merge-your-commits