test(multiple-subtitles): [PKG-MULTI-SUB-09] isolate old-core source seeking from caption loss

refactor/changes/2026-09-14-PKG-MULTI-SUB-09-switch-order.md

@@ -0,0 +1,81 @@
+# PKG-MULTI-SUB-09: distinguish source seeking from caption loss
+
+Baseline: efed847cdfcbe5110e88b31527c46e622bcc83ee. This task closes the
+misattributed switching finding with independent native evidence and corrects
+integration sequencing. No production code, public type, dependency, version,
+lockfile or built artifact changes. PKG-MULTI-SUB-05 remains doing.
+
+## Root cause evidence
+
+The original WebKit failures retained a valid Translation cue at1..3 seconds,
+but the actual media clock was approximately0.002 seconds, with no active cues.
+The subtitle was correctly absent at that time; the intended seek to1.5 had not
+landed. Those reports predate currentTime instrumentation.
+
+`multiple-subtitles-switch.spec.js` compares no plugin, current plugin, and actual
+npm plugin1.2.0 on cores5.3.0/5.4.0/candidate. Each case plays real local media,
+enters/exits native fullscreen, then performs three source changes. It captures
+native currentTime writes and media events, Promise settlement state, first seek
+completion and a second seek only if the first missed after native seeking ended.
+The property probe forwards to the original native getter/setter; it does not
+intercept or replace source operations. Outcomes and full traces are stored in
+[multiple-subtitles-switch-validation.json](../baselines/multiple-subtitles-switch-validation.json).
+
+The first WebKit run confirms5 misses in18 old-core source changes, including
+both no-plugin old-core controls. All9 candidate-core changes land correctly.
+Old core promises sometimes resolve while video.seeking is still true, allowing
+the caller's new seek to overlap native restoration. After the restoration has
+ended, a fresh seek reaches the cue interval and both old/new plugins display it.
+Candidate source/switch.ts already waits for native seeking to finish and uses
+the position restoration logic covered by BASE-LIFE-44/CORE-19. No plugin-level
+seek retry, clock rewrite, blanket source reload or global core patch is needed.
+
+The actual-main three-engine trace run has no missed first seeks in81 changes,
+but old WebKit hosts still settle during native seeking in6 of18 changes. This
+variation is retained; it is an intermittent old-host race, not proof that the
+initial misses were fixed by rerunning. All27 candidate-core changes in that run
+have seeking=false at settlement and honor the first caller seek.
+
+## Test sequencing and validation
+
+The caption combination test now waits for video.seeking=false after switchUrl
+before testing the separate user seek/caption behavior. The immediate old-call
+path is retained in the independent ordering test above, including its misses.
+This wait is a predicate on real native state, not a fixed delay or a silent
+retry of the acceptance assertion. It does not claim old switchUrl promises have
+been changed or all historical playback contracts are correct.
+
+One later full run exposed an unrelated fixture sequence: it started exiting
+fullscreen when the document property changed but before the public enter event
+arrived, resulting in old-core notifications[false,false]. The normal enter/exit
+scenario now waits for the entering event before requesting exit; the original
+strict final[true,false] assertion remains. The failed trace is retained. No
+production fullscreen correction is claimed by this test-sequencing adjustment.
+
+- Initial source WebKit ordering probe:9 cases pass their observation/candidate
+  assertions, including the5 old first-seek misses noted above.
+- Actual core and plugin main artifacts, three engines:84 cases,83 passing
+  assertions,1 fullscreen-sequencing failure. All27 ordering cases pass, and the
+  run also includes explicitly labelled historical overlap/offset defects.
+- After the entering-event predicate correction, the affected timed-combination
+  matrix is rerun unchanged otherwise:9/9 pass across three cores and engines.
+  Do not describe this as a fresh84-case all-green run.
+- The two changed browser scripts pass lint;50 CI regression tests pass. Final
+  plan/risk and risk-validator checks cover the completed task state.
+
+MULTI-SUB-SWITCH-01 is resolved as a classified old-core seeking issue already
+handled in the candidate core, with the original calls and failures preserved.
+This does not resolve MULTI-SUB-MERGE-01 or complete device/distribution acceptance.
+Use the actual loaded core/plugin hashes when comparing future runs; the fixture
+supports both source builds and the existing explicit artifact map.
+
+## Maintenance and rollback
+
+Keep source readiness, actual seek landing and caption rendering distinct when
+debugging media failures. Run the ordering probe with/without the plugin before
+adding plugin retries for an underlying player clock problem. Keep candidate
+first-seek assertions strict; old observations must never count as correct calls.
+Both suites run under yarn test:browser. A normal core/plugin build was not needed
+because production bytes are unchanged; the checked artifacts came from the prior
+completed implementation commits. Revert this independent task09 commit to remove
+the new probe and integration predicates. Local commit only, no push/publication.