Show unsaved-edit dots on manuscript issue tabs

[atomantic]

Summary

Manuscript sections persist onBlur only, so editing issue 3 and tabbing to issue 7 left the pending text with no cue anywhere on screen — the tabs showed open-note counts but nothing about unsaved state.

The saved-content baselines the beforeunload guard already tracked move from a bare ref to state (with the ref kept as an eagerly-updated mirror for the handlers that read it outside the render cycle), so ManuscriptIssueTabs can subscribe. An issue whose live content has drifted from its last server-confirmed save now renders a dot with an "unsaved edits" tooltip; it clears when the save lands and stays put when the save fails.

Making the state visible surfaced that it was also being lost. Review found three related defects on the save path, all fixed here:

1. A format switch discarded unsaved edits. Switching replaces sections and their baselines wholesale, so an edit that hadn't blurred — or whose save had failed — was dropped, and switching back refetched the server's older text. changeView now flushes pending saves first and aborts (keeping the format, the text, and the dot) if one fails.
2. Two callers could double-PATCH the same text. The blur save and the switch's flush are both fired by one click on the format switcher. Awaiting the in-flight save woke both with a stale baseline, each issuing its own PATCH and snapshotting a redundant version. Saves now chain per section, with the baseline re-check inside the link, so a link whose text the prior link already persisted no-ops.
3. The flush skipped an already-queued save. A section can be mid-save while its live text matches the baseline (type, blur, type back) — that reads as clean, so the switch went through with the save in flight and its result landed on the incoming format's save badge (saveState is keyed by issue id, not format). The flush now drains the queue first and re-derives the dirty set afterwards, so a queued save that failed is retried rather than silently passing the gate.

Two smaller things fell out of the shared isSectionDirty predicate: baseline-key construction is one baselineKey() helper instead of four inline template literals, and the predicate gates on the key being seeded — an un-seeded key means "we don't know the saved text yet", not "unsaved". The beforeunload guard inherits that.

Closes #3399

Test plan

• client: full suite green — 532 files / 6115 tests (NODE_ENV=test npx vitest run)
• New ManuscriptIssueTabs.test.jsx (6 tests): deep-linkable tabs, open-note badge, dot on only the dirty issues, no dot when clean, absent-prop tolerance, empty-sections null render
• New page-level tests in PipelineManuscriptEditor.test.jsx (7): typing marks the tab → blur-save clears it; a failed save leaves it marked; an unblurred edit is flushed before a format switch; the blur/flush race PATCHes once; text edited during an in-flight save PATCHes once; a save in flight under baseline-matching text blocks the switch until it settles; a failed save aborts the switch
• Each new test was verified to FAIL against the pre-fix code (checked individually per mechanism), so none are vacuous
• ESLint clean on all changed files