Count unreadable records in three, not one

[UnbreakableMJ]

unknown_record counted three unrelated things: an unrecognized record type, a line that is not valid JSON, and a conversation record with no uuid. Only the first is what the counter is documented to signal.

Why this matters

The merge made torn lines indistinguishable from format changes. A session that read a transcript while its harness was still appending reported 56 unknown records — every one a partial line that was complete minutes later, with nothing about the format having moved. That report reached a user three times as "the transcript format may have changed."

Same transcript, after this change:

{"unknown_record": 0, "torn_line": 1, "missing_uuid": 0}

A signal that is wrong two times in three stops being read, which costs precisely the early warning the counter exists to give. It has earned its keep before — it surfaced three miscategorized Codex tool types — and that is worth protecting.

The split

counter	meaning	what to do
unknown_record	unrecognized record type	act on it — format moved, extend an allowlist
torn_line	not valid JSON; interrupted write	nothing to fix here; often transient
missing_uuid	conversation record with no uuid	the only one where a real turn was lost

Both readers report torn lines to the new counter. Only the Claude Code reader can produce missing_uuid — the Codex reader keys on record index, not uuid.

Docs

• plugins/engram/commands/save-chat.md told the reader that any non-zero count meant the format had changed. That instruction is how the misreport reached a user; it now names all three counters and what each implies.
• CLAUDE.md documented the merge as a deliberate decision ("it correctly flags genuinely torn lines"), so it is amended rather than left contradicting the code.

Compatibility

Breaking for envelope consumers. filtered.unknown_record returns a smaller number for the same transcript; the three counters sum to the old value. Noted in CHANGELOG.md under Unreleased.

Verification

cargo fmt --check, cargo clippy --all-targets -D warnings, 245 tests (181 unit + 64 integration) — all pass. Three tests cover the new behaviour, including one per reader asserting a torn line does not register as a format change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FCJbPm4FWS99y9utbUe13z

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.