Autonomous sessions must paste raw concern-id slugs into clean-result prose — Lens 14 vs the no-opaque-codes rule

Goal

Reconcile two workflow rules that currently contradict each other for an autonomous session parking a clean-result with open non-BLOCKER concerns, and write down the sanctioned acknowledgment shape so sessions stop deriving it.

The contradiction

verify_task_body.py check "concerns audit (Lens 14)" accepts exactly two acknowledgment mechanisms for an open binding concern:

1. the LITERAL kebab-case concern_id appearing in the ## Takeaways bullets or inside a ## Results ### <result> body (the check is a plain cid in tldr_h3_text substring test); or
2. a <!-- concern-deferred: <id> --> HTML comment — which the same check REJECTS as "fabricated" unless a matching deferred event exists in concerns.jsonl.

The only writer of that event is task.py defer-concern, which is USER-ONLY: the CLI refuses without --by user (or --by reconciler for ensemble severity-downgrades), and the library rejects again defense-in-depth.

So in an autonomous session (EPM_AUTONOMOUS_SESSION=1) with no user in the loop, mechanism 2 is unavailable — correctly, since using it would manufacture consent. Mechanism 1 is therefore mandatory, and it forces raw slugs like dwfleet-align-units-key-omits-adapter-and-ft-sidecar-content into the clean-result's reader-facing prose.

That collides head-on with two standing rules:

• No opaque condition codes in clean-result prose (feedback_no_opaque_condition_codes memory; enforced in spirit by verify_task_body.py's own opaque-config-code check on figure text). A 60-character hyphenated internal slug is exactly the class of token that rule exists to keep out of a write-up.
• Dense results get a plain-language reading (CLAUDE.md): every term gets its meaning at first mention. A bare concern id has no plain reading at all.

What #2569 hit and what it did

/issue 2569 parked with 13 open binding CONCERN-severity concerns. Four were genuinely science-affecting and were folded into the relevant Results prose as stated caveats, then recorded with task.py address-concern — that path is clean and needs no user.

The other nine were caching/provenance-pin defects that change no reported number. The session first wrote <!-- concern-deferred: ... --> markers for them; the verifier correctly refused all nine as fabricated. The session declined to pass --by user (that would fabricate consent it had not received) and instead:

• wrote ONE plain-English paragraph in ## Results describing the nine in reader-legible terms (what class of defect, what it does and does not affect), explicitly framed as OPEN and unaddressed rather than deferred;
• put the nine literal ids in a collapsed <details> block inside the same section, which satisfies the substring test because <details> bodies are part of the section text — while keeping the slug soup out of the rendered reading flow;
• discovered that the <details> block then needed its own cherry-picked / random-sample disclosure in the prelude prose (the sample-block checks treat any <details> as a sample block), so the prelude had to state that the block is the complete set of nine, not a sample.

That combination PASSES the gate and reads acceptably, but every step of it was derived live against verifier feedback. It is not written down anywhere.

Acceptance criteria

1. The sanctioned autonomous acknowledgment shape is documented where a session composing a clean-result will find it — .claude/skills/clean-results/SPEC.md § the concerns/Lens-14 area, and/or .claude/agents/analyzer.md. Include the <details> + complete-set-disclosure detail, since the sample-block checks fire on it.
2. State explicitly that an autonomous session MUST NOT pass --by user to defer-concern, and that fabricating a deferral marker without the ledger event is a verifier FAIL by design (not a bug to route around).
3. Reconcile the no-opaque-codes rule with the literal-id requirement in one place, so the two are not read as contradictory: name the <details> (or equivalent) carve-out as the resolution.
4. Consider whether Lens 14 should accept a THIRD mechanism for the autonomous case — e.g. an epm:concern-acknowledged marker keyed to the id, letting the body carry only plain English while the ledger carries the machine link. Record the decision either way; a deliberate "no, keep the id in the body" is acceptable, leaving it undecided is not.
5. Whatever ships is pinned by a test.

Provenance

Surfaced by /issue 2569 (2026-08-27) while gating its clean-result body; the session's own [epm-inline-fallback] marker on #2569 records the same episode from the review-independence angle. Related: #1089 (stale deferral markers), #1891, workflow.yaml § concerns_protocol.