test(PKG-HLS-SDK-01): checkpoint direct HTTP teardown crashes

refactor/changes/2026-09-14-PKG-HLS-SDK-01-http-teardown.md

@@ -0,0 +1,70 @@
+# PKG-HLS-SDK-01: direct HTTP teardown diagnostic checkpoint
+
+Status remains doing. Baseline fe5534ac2429ac67702ca9d4121b4ab8603cb278.
+No production source, dependency, fixture, build output or acceptance assertion
+was changed. This checkpoint advances attribution, not completion.
+
+## Controls and results
+
+All five runs use frozen Hls.js1.5.17, Windows Firefox155.0, local ordinary HTTP,
+direct native video, no ArtPlayer/plugin, no route interception, no Worker
+observer and no SDK logger. Native worker output is not substituted. Browser
+process DEBUG logging is enabled and saved for every run. Existing trace capture
+and media/SDK event recording remain enabled, so these are not uninstrumented runs.
+
+| Run | Teardown | Worker | Extra pre-destroy snapshot | Pass | Fail |
+| --- | --- | --- | --- | --- | --- |
+| Original runner | reset first | yes | no | 9 | 1 |
+| Original runner control | reset first | no | no | 10 | 0 |
+| Phase/snapshot runner | reset first | yes | yes | 5 | 0 |
+| Final phase runner | reset first | yes | no | 4 | 1 |
+| Final phase runner | SDK first | yes | no | 3 | 2 |
+
+35 diagnostic iterations total:31 pass,4 fail. These are finite controls, not
+release-matrix passes or a claim about long-term failure probability. The no-worker
+control cannot substitute for required worker acceptance. Snapshot success may
+reflect changed timing; it is not evidence of a repair.
+
+The three failures in the final phase runner occur during final-state collection,
+after destroy and the immediate post-destroy state call returned. The first failure
+predates phase recording and cannot be assigned the same precise stage. Original
+errors, trace hashes and logs are retained separately from previous diagnostics.
+Both reset-first and SDK-first reproduce failures. Do not change the public core
+destroy order or callers' SDK configuration based on the previously passing
+SDK-first samples. No grouped-playback timeout occurred in these direct controls;
+the independent HLS-PLAYBACK-01 failure remains unclosed.
+
+## Browser evidence limits
+
+The installed Firefox1543 archive's TargetRegistry.js observes
+`oop-frameloader-crashed`, emits the target crash event and disposes that target.
+The logged stack reaches this observer. The source member hash and excerpt are
+recorded with the actual binary archive hash, not inferred from another release.
+This supports a browser out-of-process frame-loader crash notification; it is not
+a native stack, minidump or proof of the responsible decoder/worker/SDK component.
+The browser parent exits normally after explicit teardown. Its exit code0 does
+not negate page crashes. The SWGL framebuffer startup warning occurs in both the
+failing worker run and passing no-worker control, so it is not a root-cause claim.
+
+## Maintained diagnostic changes
+
+The runner now records host-side operation phases and the phase/time of crash
+events. Errors during final state, tracing or context close remain failures even
+after operation assertions passed. This introduces no page calls in default mode.
+`--capture-before-destroy` optionally saves media, SDK and event state to Node
+before destroying the page; it defaults off and is explicitly listed in reports.
+That extra evaluate may affect timing, which is why its passing control is kept
+separate. Existing7-second waits, retries, media, SDK and worker settings are intact.
+
+Script lint and the frozen SDK/worker archive check pass. Package source and all
+published build outputs remain unchanged, so no package rebuild is required.
+See [evidence](../baselines/hls-http-teardown-validation.json) for exact runner,
+SDK/media inputs, reports, traces and failure classification.
+
+## Next evidence
+
+HLS-CRASH-01 and HLS-PLAYBACK-01 remain open; task SDK-01 is not complete. Obtain
+native Firefox content-process crash diagnostics or a minimized MSE reproduction
+and independently investigate the grouped-stream stall. Continue other package
+tasks while keeping this gate visible. Reverting this checkpoint removes diagnostic
+phase/snapshot metadata only and restores no production behavior.