Add an annotations field to the target contract for adapter-reported checks

[ChelseaKR]

What

The target contract carries text, citations, context_ids, refused, and escalated. The real-target adapters under real_targets/ need to report two more things: how many claims the target withheld, and what the harness's own quote check found. Today both travel as bracketed suffixes on the observed text and as counters in the provenance block, which docs/real-targets.md calls legible but not a contract, and names as the obvious next change.

Add an optional annotations list to TargetResponse and to the HTTP response body: {"kind": str, "detail": str, "count": int | null, "ref": str | null} with a closed enum of kinds (withheld_claim, quote_verified, quote_not_found, quote_unverifiable, adapter_note). The evidence pack renders them per case and totals them per gate.

Why it matters

A reviewer reading a pack cannot currently see per-quote check outcomes without opening the recording beside it. Making annotations part of the versioned structure keeps the pack self-describing, keeps text as exactly what the target said, and gives issue #31's recordings a field to replay. Gates keep validating only the fields they check today; annotations inform, they never infer.

Scope

• annotations on TargetResponse, to_dict, response_from_payload (strict: unknown kind or malformed entry is a TargetProtocolError, exit 2).
• Results and pack schema version bump; a rendered "Adapter annotations" section in both forms.
• Migrate the three adapters off text suffixes; recordings carry the field.
• A target that omits the field behaves exactly as before.

Out of scope

• Any gate changing its verdict on an annotation.
• Judge rubric inputs.

Done when

• An HTTP target returning annotations renders them in evidence.md and evidence.json.
• A payload with an unknown kind is refused with exit 2 and no results file.
• The committed real-target packs regenerate with no bracketed suffix left in any observed text.
• A target omitting annotations yields a byte-identical pack to today.

Pointers

• src/gauntlet/targets.py, src/gauntlet/results.py, src/gauntlet/report.py
• real_targets/quotecheck.py, real_targets/narration.py, real_targets/rawlog.py
• docs/real-targets.md ("What the contract could and could not express"), New packs must record quote-check outcomes so a replay verifies rather than skips #31

Proposed with AI assistance.

[ChelseaKR]

Triage against origin/main (8ae3fa2): unbuilt, premise intact, automatable with no owner decisions.

Verified rather than assumed:

• TargetResponse in src/gauntlet/targets.py carries exactly text, citations, context_ids, refused, escalated — no annotations. ✔
• No open or merged PR touches it (checked all open PRs and the last 60 merged; origin carries only main).
• docs/real-targets.md still describes the bracketed-suffix workaround as legible-but-not-a-contract.

Assessment: buildable. The scope is closed, the enum is closed, and "a target that omits the field behaves exactly as before" is mechanically testable against the committed packs, which tests/test_real_target_packs.py already byte-compares.

The one hazard worth naming before someone starts. "The committed real-target packs regenerate with no bracketed suffix left in any observed text" cannot be met by re-rendering: observed comes from the recording, and the recordings under real_targets/*/results/*-raw.jsonl hold what the target said in 2026-08. Rewriting a recorded observed string to move a suffix into a structured field would be editing a recording — the thing RECORDINGS_PREDATING_QUOTE_CHECK_OUTCOMES exists to refuse. Either that bullet means "packs from new runs", or it needs fresh runs against the three targets, which costs live requests against the 100/day cap. Worth settling in the issue before code is written.

Not started by me; I took #50 this pass. Left open.

Triaged with AI assistance; premises checked against origin/main.

[ChelseaKR]

Sizing note before anyone starts this, with one item measured rather than estimated.

The Scope line "Results and pack schema version bump" is the expensive half, and it is
expensive in a specific way worth knowing up front.

Measured today on a clean main (bump RESULTS_SCHEMA_VERSION 1 → 2, run the suite,
restore, confirm green):

23 tests fail
  14  tests/test_real_target_packs.py
   7  tests/test_integrity.py
   1  tests/test_cli.py
   1  tests/test_results.py

The 14 are the mechanism: load_run_dict refuses any schema_version it does not equal,
and tests/test_real_target_packs.py loads all seven committed packs under
real_targets/*/results/ — dated records of live runs against three real third-party
systems on 2026-08-21 and 2026-08-22. They are not regenerable from this repository at
will; that is the point of committing them.

So the bump is not a version constant, it is a migration, and it needs a decision this
issue does not make: whether a bumped load_run_dict accepts version 1 as a readable
historical form (and what a v1 pack renders as in a v2 renderer), or whether the seven
packs are rewritten in place — which edits a record of what a real system said on a date.
Either answer is defensible; picking one silently is not.

Two smaller notes from reading around it:

• The other half of the "Done when" — "the committed real-target packs regenerate with no
  bracketed suffix left in any observed text" — needs the same call, since observed is
  what the target said and rewriting it retroactively is the same kind of edit.
• Add a quote_check gate that verifies cited quotes offline from a source snapshot #41 (quote_check) leans on this issue for the quotes transport. If the migration
  question above stalls, Add a quote_check gate that verifies cited quotes offline from a source snapshot #41's snapshot half (gauntlet snapshot, the manifest, the
  offline gate) is separable and does not obviously need a results-schema change — worth
  splitting rather than blocking both on one decision.

I did not start either issue for this reason. Nothing in the repository was changed by
this measurement; the tree is as it was.

Prepared with AI assistance; reviewed before submission.