verify_task_body: mechanical verbatim-sample-vs-manifest grep — FAIL fabricated '(verbatim)' sample rows

Goal

Add a mechanical check that catches fabricated "(verbatim)" sample rows in clean-result bodies: when a v4 ## Methodology Sample block labels quoted rows "(verbatim)" AND the body links a pinned bank/manifest artifact (a committed JSON/JSONL the samples are drawn from), grep each quoted sample stem into the linked manifest and FAIL on a miss.

Provenance

Surfaced by the interpretation-critic on task #2617 round 1 (2026-08-27, epm:interp-critique v1): the analyzer's Methodology sample block labeled "(verbatim)" contained a fabricated row — the bank's obj_flip_00 item text was misquoted with a substituted noun that exists nowhere in the pinned bank manifest and instead echoes the originating prompt. Same failure family as #657 (fabricated persona sample). The other three quoted samples verified. Both incidents share the shape: a quoted "verbatim" sample that a one-line grep against the linked artifact would have failed.

Sketch (target files — implementer to confirm placement)

• scripts/verify_task_body.py (preferred: a new check alongside the existing sample-block checks) or the interpretation-critic mechanical pre-pass.
• Trigger: a Sample-data slot containing "(verbatim)" + a resolvable committed artifact link in the same section.
• Behavior: extract quoted stems (quoted strings / table rows above a length floor), search the linked artifact file(s); any stem absent from every linked artifact ⇒ FAIL naming the stem and the artifact searched. Bounded: skip when the linked artifact is remote-only (HF URL without a local committed copy) — WARN instead.
• Tests: fixture with a genuine verbatim row (PASS) + a mutated row (FAIL) + remote-only link (WARN).

Acceptance

• Check fires on the #2617 round-1 shape (mutated noun in a quoted stem) and on a reconstruction of the #657 shape.
• No false FAIL on ellipsized/subset-disclosed samples that carry an explicit truncation marker.