Auditable memory (memory3l)

Run the invariant checks over one episode and return the violations. This is the tool to call to answer 'did this memory lose anything?'. I1 = a fact in the ledger no longer reaches any summary (silent loss); I2 = a slot's newest value is not visible at the top level, so the model is never told it; I3 = a fact's evidence no longer resolves to stored dialogue; I4 = extraction completeness against the raw turns; I5 = an erasure is verifiable and its evidence is gone. `ok: true` with no violations means none of the checked failure modes is present -- report that as the result, not as a guarantee about meanings.

Aggregate the invariant checks over every episode that has a ledger, including the `silent_loss_rate`. Use this for an overall health answer; use `audit` when you already know the episode and want the detail.

Return the derived current-value registry for an episode: the newest value of each attribute, with the summary it came from. This is what the model should be told at the top level. If a value the conversation established is missing here, that is an I2 failure -- confirm it with `audit`.

Return the raw user/agent turns a fact was extracted from, resolved through its evidence pointers. Use this to justify a stored value with the actual conversation rather than the summary's paraphrase. `resolved: false` means at least one evidence pointer no longer resolves -- an I3 provenance failure; that matters and should be reported, not hidden.

Given a `fact_id`, report the fact's full record and where it now lives: `live` (still in an active summary), `archived` (moved to the permanent archive, reachable by id) or `MISSING` (in the ledger but unreachable, which is an I1 silent loss). Also returns `reason` and `superseded_by`. Use this when asked why a piece of information is no longer in context.

Return one slot's value over time, oldest first, bounded by `upto_turn` when given. Use this for 'what was it before?' and for any question about change; `current` only answers 'what is it now?'. A slot is a lower-case attribute name such as `工位` or `office` -- call `current` first if you do not know the exact spelling, since this matches the slot literally.

List the episode ids that have an audit ledger. Every other tool needs one of these ids, so call this when you do not already know the episode.

Report the database this server reads, whether it exists, its size, and how many episodes carry an audit ledger. Call this first when a query returns nothing: an empty or wrong path is the usual cause, not an absent feature.

Return the episode's temporal (bitemporal-style) projection: row count, whether it is consistent with the ledger, any anomalies, and the DDL for recreating the table. Use this when memory state is projected into an external store and you need to know whether the two still agree.

List the compliance tombstones for an episode: facts deliberately erased, with their evidence pointers, which of those are still readable, and how many residual prose mentions survive. This is the deletion-verification view (I5). A non-empty `residual_prose_mentions` or a readable evidence pointer means the content is not fully gone from the store -- report that plainly instead of calling the erasure complete.