Publish the memory footer once per session bind - #12270

Summary

• At session bind the memory footer had two independent publishers of the same setStatus("memory", …) key: the initial-status pass started from bind() and the fingerprint-gated footer-live refresh after the completion drain. Whichever git read returned first decided whether one or two requests reached the engine UI stream, so the same input produced different extension UI request streams across runs.
• This change gives the bind-time publish a single owner. The footer-live segment path publishes; the initial-status pass keeps only its advisory duties (pressure dream and notify) and never writes the footer. A fresh bind's syncActive also no longer emits a restore line for a segment that was never animating. The published text is byte-identical to the previous refreshMemoryStatus output, so the footer looks the same; only the duplicate publish is gone.

Changes

• packages/omo-senpi/src/components/memory/wiring-reflection-live.ts: refreshInitialStatus now calls refreshMemoryStatus with showFooter: false and reuses the injected git repo, so a bind produces exactly one setStatus("memory", …) call regardless of git read timing. now and createGitRepo are forwarded to the footer-live wiring so both paths can be driven through one repo.
• packages/omo-senpi/src/components/memory/status-live.ts: syncActive returns early when nothing was animating (activeIdentity === undefined), so a fresh bind no longer publishes a spurious restore line. The restore path uses the captured identity/ui instead of falling back to the incoming values.
• packages/omo-senpi/src/components/memory/wiring-types.ts: adds the optional createGitRepo seam to MemoryWiringOptions.
• packages/omo-senpi/src/components/memory/wiring-reflection-live.test.ts: new describe block "bind-time memory footer publish ownership (Memory footer publishes its status a varying number of times at session bind (same input, different engine event stream) #12253)" with three co-located given/when/then cases: initial-status reads finish after the footer refresh, initial-status reads finish first, and unreadable head / unresolvable commit age. All assert the exact recorded setStatus calls.
• changes/unreleased/12253.md: ### Fixed fragment describing the behaviour change and linking the pull request.
• No changes to the committed senpi plugin bundle or goldens.

QA & Evidence

• What was tested: The new regression cases in wiring-reflection-live.test.ts drive a bind through a gated createGitRepo: one run holds the initial-status head() read behind a promise while the footer-live refresh completes, the other lets it finish first. Both observe the recorded ui.setStatus calls for the memory key. The senpi-qa skill was run with the evidence dir resolved via resolve-evidence-dir.mjs to .omo/evidence/omo-senpi-adapter/20261103-memory-footer-single-publish (gitignored, not committed).
  Observed result: RED (before the change): the gated case failed because the memory key was published twice — expect(calls).toEqual([{ key: MEMORY_STATUS_KEY, text: expectedText }]) received a second identical entry once the held head() read resolved. GREEN (after the change): the three #given/#when/#then cases under "bind-time memory footer publish ownership (Memory footer publishes its status a varying number of times at session bind (same input, different engine event stream) #12253)" pass with 0 failures, and the existing cases in the file still pass; the unreadable-head and unresolvable-age cases publish nothing.
  Artifact: drive-self-test.txt, drive-live.json, README.md under the evidence dir above (sanitized: no tokens, auth headers or credentials).
  Why sufficient: The cases pin the exact UI request stream rather than the rendered text, which is what diverged between runs, and they cover both interleavings of the two git reads plus the no-data paths. The live artifact records the real engine run alongside the self-test.

sha256sum .omo/evidence/omo-senpi-adapter/20261103-memory-footer-single-publish/drive-self-test.txt
sha256sum .omo/evidence/omo-senpi-adapter/20261103-memory-footer-single-publish/drive-live.json
sha256sum .omo/evidence/omo-senpi-adapter/20261103-memory-footer-single-publish/README.md

Digests are recorded in the sanitized README.md of the evidence dir; the dir is gitignored so the artifacts are not attached to the branch.

Risks & Residuals

• The initial-status pass no longer publishes the footer, so if the footer-live path cannot resolve a head or commit age, the bind publishes no footer line at all. That is the intended deterministic behaviour and is covered by the third test.
• The desktop equivalence cell E17 golden is not re-recorded here; this change only removes the nondeterminism that blocked it.
• lint:ratchet baseline left unchanged: the changed files add no new findings and this change does not remove tracked debt.
• CLA not signed by a human; it must be signed before the contribution is acceptable, and nothing in this pull request implies otherwise.
• Fork-target PR: no golden or bundle regeneration, no dependency changes, no reformatting outside the touched files.

Screenshots

Not applicable: no visible UI change. The footer text is byte-identical to the previous output; only the duplicate publish is removed.

Automated Checks

bun run typecheck
tsgo --noEmit -p packages/omo-senpi/tsconfig.json
bun run test:senpi
bun test
bun run lint:ratchet

All pass locally on Bun 1.4.4 / Node 24.

Related Issues

• Memory footer publishes its status a varying number of times at session bind (same input, different engine event stream) #12253

Summary by cubic

Fixes the session-bind memory footer so each bind publishes exactly one footer status, and a later session binding to the same git state still gets its own footer.

• The footer-live segment path owns the publish; the initial-status pass keeps only pressure dream and notify.
• The publish runs after the completion-drain try, so a drain failure no longer drops the footer.
• syncActive clears the shared dedupe cache on a fresh bind without emitting a restore line, so a second session bound to identical git state still publishes.
• Adds regression tests for both git read interleavings, the unreadable-head / unresolvable-age paths, and the second-session case; published footer text stays byte-identical.

^{Written for commit e6e6135. Summary will update on new commits.}