fix(agent): poll convergence tears down a deleted def's live agent (orphan fix)

[unforced]

What

Deleting an #agent/definition note out-of-band (anything other than the admin-SPA delete button — a direct vault API/MCP delete, an Obsidian sync, another surface) left the live agent + wake channel running until the next daemon restart. The orphaned agent kept answering messages after its def was gone from the vault.

Root cause

There is no vault deleted trigger. The hub's connection engine maps only note.created/note.updated to vault-trigger verbs (parachute-hub admin-connections eventToVaultEvents — note.deleted has no mapping and throws). So a delete never fires the reactive reload(...,"deleted") teardown.

That leaves the 60s poll as the only automatic convergence path for a deletion — but loadAll → pruneRemovedDefs only pruned the removed def's grants (reconcileGrants(agent, [])). It never called deregisterByNote, so the live programmatic registration + the wake channel survived. Only the admin-SPA delete button worked, because deleteDef does the teardown in-process; every other delete path orphaned the agent.

Fix

• pruneRemovedDefs now does the full teardown the reactive confirmedRemoval already does: deregisterByNote (drop the programmatic registration + wake channel) then prune grants.
• Truncation guard (because the diff is now destructive): a def list at the page cap (DEF_LIST_LIMIT = 500) is treated as possibly-partial — not a confident set — so a truncated read can't be mistaken for deletions and tear down live agents. Sits alongside the existing list-failure guard. This also hardens the pre-existing grant-prune, which had the same latent exposure.
• Corrected the loadAll/pruneRemovedDefs/daemon-boot comments that wrongly claimed the poll already converged deletes.

How it was found

Live end-to-end verification of the agent-to-agent callback feature (#123). While cleaning up two throwaway test agents, deleting their def notes did not deregister them (/health still listed them 72s later, past the 60s poll). The reactive deleted webhook never fired → traced to the missing vault delete-trigger → the poll's grant-only convergence. (The callback feature itself verified clean: worker→callback→orchestrator chain with the full metadata contract, no loop.)

Tests

• loadAll tears down a removed def — deregister + removeChannel via the poll (the no-delete-trigger regression).
• loadAll skips the removed-def teardown on a truncated (page-cap) read — no spurious deregister.
• The existing agent-grants: grant rows orphaned when an #agent/definition note is deleted #96 removal test now also asserts teardown.
• Gates: tsc --noEmit clean; bun test ./src 1019 pass / 0 fail; SPA (vitest) 110 / 110.

Note for the merger

origin/ag-unforced-dev currently carries two unmerged commits with no open PR (8aca779, 9db9689 — a "robustness hardening (PR #3)" series, looks like another session's WIP). To avoid clobbering/entangling that branch, this fix is isolated on a one-off branch off main. It is independent: that series leaves pruneRemovedDefs unchanged (still grant-only), so there's no semantic overlap — at most a trivial textual conflict in agent-defs.ts whichever lands second.

Per RC versioning, no version bump in this PR (bump + tag on ship).

[unforced]

Folded the two diff-relevant reviewer nits inline (non-blocking):

• Nit 1 — the truncation test now does a third confident pass proving the seen-set survived the truncated pass (researcher is still caught as a real removal afterward), so it's a complete proof of the invariant rather than a partial one.
• Nit 2 — added an inline note at the notes.length < DEF_LIST_LIMIT guard boundary so the off-by-one reads unambiguously.

Nit 3 (seen-set clearing relies on rebuildSeenDefs running right after) is already acknowledged in the pruneRemovedDefs comment — left as-is since adding a redundant .delete() would just duplicate work the rebuild does.

Gates after folding: typecheck clean; bun test ./src/agent-defs.test.ts 70 pass / 0 fail (207 assertions).