fix: citation numbering gaps, flag-click annotation reset, stale client directory

[rprabhat]

Summary

Items #1, #2 (re-scoped), #3 from the consolidated fix plan — bundled because each is the same failure class: the product shows the user something that doesn't add up.

• Answer transparency & explainability #3 - Citation numbering gap. _parse_citations only ever returned entries for numbers the model happened to cite ([1][2][4][5] stayed [1][2][4][5]). New _renumber_citations remaps both the rendered answer text and the citations array to a contiguous 1..N sequence in first-appearance order. Applies to newly generated answers only - existing stored conversations keep their original numbering.
• Backend audit: latency, agent cost, answer correctness & personalisation improvements #1 - Flag-click removes the underline instead of explaining it. Root cause: AnnotatableMarkdown's visibleThreads was rebuilt with a plain .filter() on every render, breaking the referential stability RecogitoLayer's setAnnotations effect depends on - triggering a full highlight-layer replace/rebuild far more often than needed, including on renders that land mid-click. Fixed by memoizing on [threads, showResolved].
• Ports-and-adapters backend refactor + LangGraph agent loop #2 - Workspace Clients directory shows 0 activity for clients with real work on file. Re-scoped mid-implementation: the original plan called for a data backfill, but a dry-run of the existing scripts/backfill_engagements.py against prd found zero unlinked buckets - the data isn't missing. Actual cause: WorkspaceClientsTable's fetch only ran once on mount, so Radix's persistent TabsContent never learned about activity created after the tab was first visited. Workspace's Tabs is now controlled and hands the table a fresh key each time the Clients tab is selected, forcing a refetch.

Verification

• Backend: 842 tests pass (4 new, covering renumbering - gap-closing, first-appearance ordering, multi-number brackets, no-op when nothing cited).
• Frontend: tsc --noEmit and eslint clean.
• Live-verified in browser (demo login, Nepean Tradie Accountants persona, the exact conversation with the citation gap the accountant review flagged): clicked two separate flagged claims - both opened the "Needs review before relying on this" popup with full detail, and the underline stayed intact after closing each popup. Confirms the flag-click fix works end-to-end, not just in isolation.

Reviewer notes

• Ports-and-adapters backend refactor + LangGraph agent loop #2's root-cause pivot (data -> code) was surfaced and confirmed with the user mid-implementation before proceeding - see commit message for the dry-run evidence.
• Config-driven model routing + optional OpenCode provider #4 (generation-failure recovery UX) and Structured JSON logging + request-id middleware #8 (bulk client import / PM-tool integration) remain out of scope for this PR per the plan - separate, larger pieces of work.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
taxflow-dashboard	Ready	Preview	Aug 8, 2026 8:00am