[bokelley]

The narrowing checker missed three published conflict vectors: pixel-density sets, a duration range escaping its upper bound, and an exact width mismatch. Seller-asserted references also passed through projection without a way to provide their requirements for checking.

Cover acceptance sets, exact numeric values, range containment and min/max boundaries, including legal unbounded duration endpoints and exact-duration precedence. Run all eleven unchanged AdCP 3.2.7 narrowing vectors. Add an optional synchronous v1_requirements_lookup to declaration/product projection so callers with a provisioned catalog can check each full format reference and receive the existing FORMAT_DECLARATION_DIVERGENT advisory at its indexed location.

Refs #1749. Projection preserves seller references. An omitted lookup or missing entry remains explicitly unchecked; the SDK does not fetch or certify a catalog. Lookup failures propagate. No schemas, version pins or generated models change.

Give the context-ID integration test fixture the unique private name _ContextIdEchoHandler. This prevents another test module’s _EchoHandler registration from hiding create_media_buy during full-suite collection. The test assertions and library behavior are unchanged.

Validation

• Full local gates in Docker: make lint, make typecheck-all, and make test passed. 18621 passed, 2729 skipped, 10 deselected, 1 xfailed, 1017 warnings in 3738.16s (1:02:18); coverage 92.14%.
• Docker: 123 focused tests passed across narrowing, projection, round trips and public API.
• The original checker failed the three published conflict vectors; new controls cover range/exact boundaries, nullable endpoints and precedence in both directions.
• Projection controls verify distinct agents sharing an ID, per-product advisory locations, missing catalog entries and lookup exceptions.
• Independent code and security reviews passed after addressing interval edge cases.
• Normal repository commit hooks run formatting, lint, source/adopter type checks, Bandit and file checks.

SDK 10 forward port: #1759.

[aao-secretariat]

Ladon verdict: Approve

Approve — additive, backward-compatible narrowing support with full test coverage.

What I checked:

• Adds interval/exact/pixel-ratio narrowing to check_narrows (narrowing.py L105-130) and an optional synchronous v1_requirements_lookup to declaration/product projection. Both are additive, keyword-only params — no schema, generated-code, or public-surface break.
• The new _RANGE_FIELDS loop was traced against the existing _MAX_FIELDS/_MIN_FIELDS/_EXACT_FIELDS loops; the add_interval_divergence dedup (narrowing.py:116-125) correctly suppresses duplicate scalar/named-cap diagnostics, and the broad not_equal suppression cannot hide a real divergence (a not_equal on width/duration_ms_exact implies both sides exact, collapsing to the same single conflict).
• Unbounded (null) endpoints, exact-wins-over-range precedence, and the max_duration_ms→duration_ms_exact fallback behave per the 11 published AdCP 3.2.7 vectors plus added boundary tests.
• v1_requirements_lookup is a caller-supplied synchronous callback whose failures deliberately propagate (documented, test-covered) — no SDK-side timeout owed.
• No adcp.* public-surface signature/type changes, no generated-code hand-edits, no import-layering violations, no CI-gate disabling.

No blocking or medium findings. No gated paths, no high-risk flag, no author-team gate (review_decision REVIEW_REQUIRED but gated_paths is false, so row 2 does not fire). Falls through to row 9.

Blocking findings: none.
Medium findings: none.

[aao-secretariat]

Ladon verdict: Approve

Approve — subsequent review of a clean diff, no blocking or medium findings.

This push is a small delta over a prior clean pass. The only behavioral-adjacent change is a test-isolation rename in tests/integration/test_a2a_context_id.py (_EchoHandler → _ContextIdEchoHandler), applied consistently across all 5 occurrences (definition + 4 call sites) with no stale references left. The rename avoids a handler-registration collision with another test module that was shadowing create_media_buy under full-suite collection — a sound, test-only fix with no change to library behavior or the public adcp.* surface.

Decision-table walkthrough:

• Rows 1 (critical/high): no findings — skip.
• Row 2 (gated_paths): gated_paths is false — skip (review_decision REVIEW_REQUIRED is immaterial here).
• Rows 3–5 (high_risk): high_risk is false — skip.
• Row 6 (sticky escalate): prior decision was approve, not escalate — skip.
• Row 7 (no-auto-approve team): no team match — skip.
• Row 8 (≥3 medium): zero medium findings — skip.
• Row 9: approve.

No semver-signal, forward-compat, import-layering, generated-code, credential, CI-gate, or version-translation concerns are implicated by this test rename.