Replay entry point, cache hit through replay, profile-aware path tests (#790 #788 #787)

.github/workflows/ci.yml

@@ -297,6 +297,12 @@ jobs:
           CI: "true"
         run: cargo test --test integration_test -- --ignored --nocapture --test-threads=1
 
+      - name: Run replay-path tests (fresh vs cache vs replay)
+        env:
+          AUDITMYSITE_BROWSER: ${{ steps.setup-chrome.outputs.chrome-path }}
+          CI: "true"
+        run: cargo test --test replay_paths_test -- --ignored --nocapture --test-threads=1
+
       # Schema/vulnerable-libs/easy-language detection-corpus tests
       # (#553-#558; the WCAG corpus has its own job below): real Chrome, diffed against ground-truth fixtures.
       # These are separate `#[ignore]`-gated test binaries and were never

CHANGELOG.md

@@ -9,6 +9,7 @@ and verification for each change is in the
 
 ### Added
 
+- **Replay:** `auditmysite replay <snapshot.json | cache directory | URL>` evaluates a stored snapshot without browser or network and reports it with `execution.environment.source: replay` and the verdict as exit code. `--session`, `--state-id` and `--storage-state` given to it are assertions about the stored measurement profile: a snapshot measured under another one is refused (`ReplayError::OtherMeasurement`, naming field, stored and asserted value) instead of being shown as the same measurement; `--dns-check`, `--check-ssr-content`, `--isolate-third-party-impact` and `--ai-transparency` are refused, because a replay shows the checks its run recorded. `--reuse-cache` answers through the same replay (`source: cache`); an unreadable, incomplete, version-mismatched or differently configured entry is a miss that prints its reason (`CacheMiss`), and the lossy reconstruction is no longer used for cache hits. `tests/replay_paths_test.rs` (Chrome, local fixture) checks that fresh run, cache hit and replay give one canonical model and that a single-page app whose content changes after a client-side navigation is explained by its measurement profile (`audit::profile_differences`).
 - **Snapshot:** `snapshot.json` now holds the whole collected report, the document identity (requested/final URL, status, redirects), the run language and the versions it was written under (snapshot schema, fingerprint, collection contract, measurement profile, evaluation, decision contract, rule-set digest). `audit::replay` evaluates it without browser or network into the same model as the fresh run; `output::format_json_replay` exports it with `detail.module_detail: full`. Incomplete or version-mismatched snapshots are reported as such (`ReplayError`) and are a cache miss, never read under current semantics. Cache entries of earlier builds are not reused. Details: [Snapshot, replay and validation paths](https://github.com/casoon/auditmysite/blob/main/docs/OUTPUT_CONTRACT.md#snapshot-replay-and-validation-paths).
 - **Validation:** `tests/validation_path_matrix.rs` checks that every fallible export entry refuses contradictory data and that unchecked `normalize()` calls in `cli`, `output` and `studio` do not grow.
 - **Navigation:** `navigation.redirect_chain` lists the main document's redirects (URL and status, in order) next to `requested_url`, `final_url` and `redirect_count`.

docs/OUTPUT_CONTRACT.md

@@ -130,7 +130,8 @@ The cache entry's `snapshot.json` is the collected evidence, independent of any
 
 - `versions` records the snapshot schema, the content fingerprint, the collection contract, the measurement-profile version, the evaluation version (normalization, scoring, decision derivation), the decision contract and a digest of the shared rule catalogue. A snapshot is replayed only when all of them equal the current values.
 - `audit::replay(snapshot)` evaluates a snapshot without browser or network and returns the same `NormalizedReport` as the fresh run. Only `timestamp`, `duration_ms` and `execution.environment.source` (`replay`) may differ. `output::format_json_replay` exports it with `detail.module_detail: full`.
-- A snapshot that lacks parts (`ReplayError::Incomplete`, naming them) or was written under other versions (`ReplayError::Incompatible`, naming field, stored and current value) is never evaluated under today's semantics. `load_artifacts` treats it as a cache miss.
+- A snapshot that lacks parts (`ReplayError::Incomplete`, naming them) or was written under other versions (`ReplayError::Incompatible`, naming field, stored and current value) is never evaluated under today's semantics. `load_artifacts` treats it as a cache miss; `lookup_artifacts` and `cached_report` return the reason (`CacheMiss`: absent, unreadable, other configuration, not replayable), which the CLI prints.
+- `--reuse-cache` answers through `replay_report_as(snapshot, "cache")`, `auditmysite replay <snapshot|directory|URL>` through `replay_report`. `SnapshotArtifact::check_measurement(&ExpectedProfile)` refuses a snapshot whose session policy, state id or storage-state digest differs from the asserted one (`ReplayError::OtherMeasurement`); `audit::profile_differences` names the differences between two profiles. A replay measures nothing, so the extra-check flags are refused for it (support matrix, mode `Replay`).
 - `format_json_cached` works on a stored `NormalizedReport` alone and says so with `module_detail: not_restored`; an empty module block there never stands for restored data.
 - Bump `SNAPSHOT_SCHEMA_VERSION` when the shape of the collected report changes in a way old data would survive deserialization, and `EVALUATION_VERSION` when a change moves a score, finding or verdict for unchanged input.