EmbeddedTransport ignores recordedSince's cursor and returns the whole journal

[EtaCassiopeia]

What

EmbeddedTransport now overrides recordedSince (delegated unconditionally) and the scoped clearRecorded(int, List<MatchClause>), routing both to the lazily-started in-process admin server — the pattern #174/#177 established for events(). The unfiltered clearRecorded(int) deliberately stays on the direct C-ABI.

Two different failures on one transport, from one missing override:

| Call | Before | After |
|---|---|
| Imposter#recordedPage / recordedSince | whole journal, since ignored, empty cursor (silent) | real paged read, real cursor |
| Space#recordedPage / recordedSince | UnsupportedOperationException (loud) | works |
| Imposter#clearRecorded(match) | UnsupportedOperationException | works, genuinely scoped |

Why recordedSince delegates unconditionally but clearRecorded does not

recordedSince has an observable return the C-ABI cannot supply — the cursor. Answering the bare recordedPage() baseline from the C-ABI would report no cursor on the first read and a real one on the second, which is exactly the signal a tail is written to treat as "this transport cannot tail". RecordedCursorIT's "0 is a cursor, not an absence" assertion fails under a conditional split.

clearRecorded is void — no observable difference — so the empty-clause case stays on the C-ABI and never starts a server it does not need. Both arms mean the same thing: an empty clause list makes the admin form emit DELETE .../savedRequests with no query, which is rift_clear_recorded's own behaviour, and one ImposterManager backs both routes.

Gate

Confirmed red first on the EMBEDDED conformance lane — 5 distinct failures, Skipped 0, one per acceptance criterion — by flipping the gates before implementing:

• EventStreamIT: deleted the EMBEDDED fork so assertEquals(page.nextIndex(), pushed.index()) holds on both lanes, and removed the needsCursor SPAWN-only gate so the reconcile-loop test runs on embedded. Both were tripwires planted by EmbeddedTransport: delegate events() to the in-process admin server instead of throwing #177 to fail the day this landed.
• RecordedCursorIT, SpaceCursorIT, MatchClauseIT: un-gated from SPAWN-only to lane.isAvailable().

MatchClauseIT is the one that matters most: it proves the whole clause grammar survives the delegation — clause intersection, header wire-encoding with a space and an =, method/path clauses, case-sensitivity, and that a filtered tail's cursor advances past entries the filter rejected. Without it, "embedded can filter" would only have been proven for single-clause cases.

EMBEDDED coverage: 7 → 12 tests. Both lanes 12/12, Skipped 0. Full mvn verify green on JDK 17 and JDK 21; javadoc:javadoc clean.

Stale claims retired

This gap was asserted as fact in nine places. All corrected: docs/events.md (reconcile warning, transport table, start-up-cost line), docs/spaces.md, docs/design/sdk-api.md (§8.1, §8.2, FFI-mapping list), and javadoc on RiftTransport.recordedSince, Imposter, Space, EmbeddedTransport, plus three "like the in-process FFI one" glosses in core tests and a doubly-stale comment in EmbeddedTransportIT.

Known follow-up

engine()/gated() scaffolding is now duplicated across four conformance ITs. Extraction is a separate refactor, filed rather than bundled here.

Closes #175