Source fidelity audit (SV2-MIG-02)

Before legacy transcript service (agentibridge) history is imported, each historical session is classified by what survives of it. The audit is code: hooks.memory.source_audit.audit_source classifies one session and hooks.memory.source_audit.SourceAudit holds the audit record across runs. It reads; it never rebuilds the legacy service, its collector or its stores.

Classes

Class	Survives	Import confidence
full_raw	the provider’s raw JSONL transcript	lossless
preview_only	the legacy Redis entry preview, no raw transcript	lossy
vector_only	legacy vector rows only	fragment
missing	nothing	none

SourceAudit.certify gives a lossless import label only to a full_raw source. A preview-only, vector-only or missing source is refused with not_lossless. SourceAudit.report lists the four classes separately and counts them in source_fidelity_classes.

From a shell, python -m hooks.memory.source_audit run STATE OBSERVATIONS REVISION audits a JSON list of observations, each raw naming a transcript file beside the list, and saves the record in STATE; certify STATE SOURCE and report STATE read it. A refusal prints its error class and detail on stderr and exits 1.

Legacy truncation measured

Read from the retired service’s store, collector and embedding code:

Legacy store	Truncation	Measured as
Redis entry list	each entry’s content cut to 500 characters	capped (raw known) or possibly_capped (preview only)
Redis entry list	list trimmed to the newest 500 entries	trimmed_head range, or possibly_trimmed
Incremental collector	a read resumed at a stored offset discards the next full line	incremental_offset_gap range
Incremental collector	entries written after the last collection	uncollected_tail range
Vector rows	text preview cut to 1000 characters	rows compared at that cap
Vector rows	chunk_idx numbered over the trimmed preview, upserted in place	renumbered rows and orphans
Any of the above	no preview entry matches the raw transcript, so a range has no anchor	unplaced range

Missing ranges are entry index ranges in the raw transcript, so they exist only for full_raw sources. A preview-only source reports missing_ranges: "unknown"; the audit never fills absent text from a preview or a vector row. Limits are LegacyLimits (500, 500, 1000 by default).

Reruns

Each classification keeps the hashes of the raw bytes, preview and vector rows it was made from. A rerun after more raw files appear upgrades only the sources whose class rose; unchanged sources replay. A lower observation, such as a raw file gone again, is reported kept and never downgrades the held classification. A run that would change the record at or below its current revision is refused as stale_revision with nothing written; a run in which every source replays is accepted as a no op. A run naming one source twice is refused as duplicate_source.

Rollback

The classification record is a projection of the sources. To roll back, delete that record and keep the package evidence files, which hold every source hash and case output. The raw files, Redis previews and vector rows are authoritative and are never deleted by the audit.