fix(mcp): smem_edit recomputes expires_at when the type changes, without reviving tombstones or ephemerals

[RobertSigmundsson]

Summary

• smem_edit now recomputes expires_at from the new type's default when the memory type changes, instead of leaving the old type's TTL in place.
• The recompute is skipped for soft-deleted and ephemeral memories, both of which carry an expiry that means something and must survive a type change.
• Adds four tests: one per direction of the recompute, one per guard.

Why

_edit's type-change branch swapped memory_type and left the old TTL untouched. DEFAULT_EXPIRY_DAYS is consulted only at creation time, through remember_handler's expires_in_days, and TypedMemory is a frozen dataclass with no __post_init__, so nothing recomputed it afterwards either.

The result was wrong in both directions. A DECISION (90 days by default) edited to FACT (no default expiry) still expired roughly ninety days out, so a fact quietly vanished on the decision's clock. Going the other way, a FACT edited to TODO or ERROR (30 days each) never acquired a finite expiry and persisted indefinitely.

The fix reads DEFAULT_EXPIRY_DAYS for the new type after the swap and either clears expires_at or calls TypedMemory.extend_expiry — a helper that already existed on the dataclass and, until now, had no caller anywhere in mcp/.

Two memories the recompute must not touch

An unconditional recompute is a data-integrity bug of its own, so it is guarded twice:

• A soft-deleted memory. _forget(hard=False) tombstones by setting expires_at=utcnow(). Recomputing from the new type's default would hand a deliberately forgotten memory an open-ended life — a FACT edit would make it immortal. The guard skips any record whose expires_at is at or before now.
• An ephemeral memory. remember_handler gives these a one-day TTL and flags the anchor neuron ephemeral=True. Clearing that expiry would turn a memory documented as "auto-expires after 24h, never synced" into a permanent one. The guard reads the anchor neuron and skips the recompute when the flag is set.

In both cases the type swap still applies; only the TTL recompute is skipped. _forget is untouched and remains the only other writer of expires_at.

Changes

• mcp/lifecycle_handler.py: import DEFAULT_EXPIRY_DAYS; after the type swap, compute the tombstone and ephemeral conditions and, when neither holds, clear or extend the expiry.
• tests/unit/test_edit_forget.py: four tests, capturing the TypedMemory handed to storage.update_typed_memory — the interface between _edit and persistence, rather than any downstream effect. test_edit_type_change_recomputes_ttl (DECISION to FACT clears it), test_edit_type_change_picks_up_finite_expiry (FACT to TODO gains 30 days), test_edit_type_change_does_not_resurrect_soft_deleted, and test_edit_type_change_preserves_ephemeral_ttl.

Not touched, on purpose

No new argument is added to smem_edit. Exposing expires_at explicitly would be the more flexible answer, but it is a schema change and a decision for you rather than for a bug fix. If you would rather have that, this PR is easy to rebase into it.

Test plan

• pytest tests/unit/test_edit_forget.py — 19 passed, against 15 on main; the delta is exactly the four tests added here.
• Differential control, run twice, because the two halves of this change need separate proof:
  • Reverting the whole hunk to main and keeping the tests: 2 failed, 17 passed — the two direction tests fail, and the two guard tests pass, because with no recompute at all there is nothing for a guard to prevent.
  • Reverting only the guards, keeping the recompute: 2 failed, 17 passed — and this time it is exactly the two guard tests. The tombstone test sees its expiry resurrected; the ephemeral test sees its expiry cleared.
  • Neither pair can pass by accident, which is the property the first version of this change lacked.
• pytest tests/ -m "not stress" -n 4 against a live SurrealDB v3.2.0 — 7287 passed, 48 skipped, 1 xfailed, which is main's 7283 plus the four tests here. Two tests in tests/unit/test_dashboard_brains_scope.py fail on this branch and on main alike: they want a live database and collide with one another under -n. Both pass when that file is run on its own.
• ruff check src/ tests/ clean; ruff format --check src/ tests/ reports 739 files already formatted.
• mypy src/ --ignore-missing-imports — success, no issues found in 354 source files.
• Coverage under the CI gate: 72.39%, against 72.36% on main.
• CHANGELOG.md untouched — left to the release entry, as with fix(storage): bind datetimes in time comparisons so they select by value #191–fix(memory): refresh content-derived fields on compress, restore, and refine #193.

Verified by

@RobertSigmundsson

[acidkill]

Reviewed the full diff. Type change recomputes expires_at from DEFAULT_EXPIRY_DAYS[new_type]; the tombstone/ephemeral carve-outs prevent resurrecting soft-deleted memories or making ephemerals immortal. Both directions tested.