feat(reporting): support per-metric source availability evidence

[bokelley]

Inline reporting fetches can now report different availability for each requested constituent/metric cell. A single constituent can retain impressions, explicitly report zero clicks, mark viewability delayed, and explain unsupported completed views in one sealed manifest.

Closes #1170

• Add immutable MetricEvidence, optional nested InlineFetchResult.cell_availability, and helpers for complete data, an unsupported metric everywhere, and a delayed metric through a watermark.
• Validate requested keys, duplicate mapping entries, status evidence, watermarks, and contradictory present/zero row assertions before staging. Explicit cells override constituent defaults; coverage and eligible control totals follow the resolved cells.
• Keep semantic-contract identity exclusively in the selected SDK offering. The manifest schema and fingerprint/replay algorithms remain unchanged.
• Add sync/async adopter type fixtures, documentation, and table-driven conformance tests.

Compatibility

Bare rows, None, existing positional InlineFetchResult arguments, and omitted/empty evidence maps retain their input behavior and constituent defaults. An independent review compared 12 representative legacy sealed shapes byte-for-byte with origin/main; all remain identical. Legacy-derived incomplete cells and unmatched staged rows continue to participate in their existing control-total behavior. A metric total is withheld only when the adapter explicitly declares a cell of that metric unavailable.

Validation

• Initial focused reporting tests, including PostgreSQL 16 ledger/lifecycle conformance: 458 passed.
• Initial full local PostgreSQL gate: 8,531 passed, 34 skipped, 9 deselected, 1 expected failure; 84.90% coverage.
• Independent review: 113 focused, 479 reporting/inline with 16 skipped, and 8,374 full-suite with 44 skipped, 9 deselected, 1 expected failure; strict typing, packaging hooks, and generated-code validation passed.
• Compatibility probes: 12-shape byte parity against origin/main, a 1,920-case exhaustive sweep, and a 10,800-case randomized validation fuzz, with no crashes or validation failures.
• GitHub CI passes Python 3.10–3.13, PostgreSQL conformance, schema validation, security, downstream imports, and JavaScript storyboard lanes.

Residual risks and limits

• Omitted cells retain legacy derivation; adapters must explicitly declare exceptions. The existing v1 wire rule still disallows mixing available and unavailable cells in a zero-row batch.
• Structural/data contradictions currently raise from execute() rather than becoming typed source failures. This widens a pre-existing class and should be handled separately at the worker boundary.
• A cell watermark later than the batch watermark promotes the batch watermark; a later publication that drops that evidence can correctly fail revision-sequence regression checks.
• Unavailable cells may retain raw staged values; their metric control total is withheld, and consumers must follow the explicit evidence.
• Live GAM/FreeWheel reads were not exercised in this PR; the closing feat(reporting): provide an adapter-first ReliableReportingService and conformance kit #1172 interop/conformance slice covers adapter integration.

---

Open workspace in Conductor

[aao-secretariat]

Ladon verdict: Approve

Approve — additive per-metric source availability evidence.

Checked:

• Change confined to src/adcp/reporting/inline_source.py with correct feat: semver signal.
• Appended dataclass field preserves positional compatibility (no public-API break, no required→optional flip).
• Fail-closed validation runs before staging.
• Comprehensive conformance (tests/conformance/reporting/test_inline_cell_availability.py) and type-check coverage added.
• No type-system layering, generated-code, forward-compat, or credential concerns touched.

Decision path: no critical/high/medium findings; gated_paths false; high_risk false; no no-auto-approve team match; review_decision: REVIEW_REQUIRED does not gate absent gated_paths. Falls through to row 9 → approve.

[aao-secretariat]

Ladon verdict: Approve

Approve

Subsequent pass to a previously-approved PR (prior head 2037b40, no findings). The delta narrows control-total withdrawal to explicitly declared incomplete cells so the total is a checksum over staged rows, restoring legacy sealed-bytes/fingerprint behavior for adopters supplying no cell_availability and closing a latent replay-fingerprint regression. Reviewer verified the Collection import, freshness-gate interaction, None-fetch path, and that no test was disabled — no public-surface or semver break.

No blocking (critical/high) findings and no medium findings.

Gate checks:

• No no-auto-approve team match (row 7 n/a).
• gated_paths: false → row 2 n/a (review_decision REVIEW_REQUIRED does not gate because paths are not gated).
• high_risk: false → rows 3/5 n/a.
• Prior decision was approve → row 6 n/a.
• Zero medium findings → row 8 n/a.

Falls through to row 9: approve.