feat(session): deterministic junk classifier for chat sessions #1303

Acceptance Criteria

• A greeting-titled session (hello/hi/howdy/test/smoke vocabulary) with < 40 user chars, no todos, ≤ 6 assistant messages classifies junk-greeting.
• A session with zero assistant messages and ≤ 1 user message classifies dead-cast, regardless of title length.
• A session with zero user and zero assistant messages classifies probe.
• A substantive session with a greeting title but > 6 assistant messages classifies substantive.
• A session with pending todos never classifies into a junk bucket.
• The classifier is a pure function of the features object — no DB access, no I/O, no network.

Testing Decisions

New vitest file in the CLI package's test dir, following the pure-core test pattern of the retention/recap modules. Table-driven: one case per acceptance criterion plus the 2026-09-20 sweep calibration cases (greeting titles, one-word reply test, fleet-smoke, live campaign session).

Key Decisions

• Bucket set is closed: junk-greeting | dead-cast | probe | substantive. Anything unrecognized defaults substantive (curation never loses work by default).
• Thresholds (40-char, ≤6 assistant turns, vocabulary list) are module constants with named exports — the nightly archiver reads them; no hidden magic numbers.
• DB-row → features mapping lives OUTSIDE the pure module (caller maps rows); keeps the classifier testable without a DB.

Constraints & Invariants

• Deterministic only — no LLM, no network in this path (the Jev middle layer is a future slice).
• Never returns junk for a session it has insufficient features for.

Prior Art

• session_retention.ts (CLI package) — the pure-core + ops-preference pattern and threshold conventions.
• session_recap.ts (extension plugin) — noise-title vocabulary to unify with.
• The open-threads skill's noise-title filter — same vocabulary, do not duplicate divergence.

Source

Part of #1301 · spec of record: personal vault specs/spec-20260920-session-curation.md