refactor(storage): isolate dialect search behind Bun capabilities

[mariusvniekerk]

Full-text and vector candidate retrieval legitimately differs by engine, but hydration, visibility, filtering, deduplication, and public pagination should not. This layer narrows dialect-specific search to capability adapters and moves the observable search pipeline into the shared Bun store.

DuckDB tool-content search now uses canonical session and message-ordinal identity instead of disposable physical row IDs. Timestamp ordering and portable search choices are supplied through narrow adapter hooks, while engine-specific relevance ranks remain allowed where the engines expose genuinely different semantics. Stack 3 of 5.

^{generated by a clanker}

[roborev-ci]

roborev: Combined Review (7bb6694)

Code review passed with no Medium, High, or Critical findings.

---

Reviewers: 2 done | Synthesis: codex, 6s | Total: 20m56s

[roborev-ci]

roborev: Combined Review (5fcc6ce)

No issues found.

---

Reviewers: 2 done | Synthesis: codex | Total: 10m38s

[roborev-ci]

roborev: Combined Review (a27f90c)

Medium-severity issue found in remote Quack regex search.

Medium

• internal/db/bun_search_content.go:686 — Direct QueryContext bypasses DuckDB’s Bun connection resolver, causing regex searches against a remote Quack store to query the local client database instead of the remote mirror. Stream rows through a Bun raw query that uses the resolver, and add Quack regex-search coverage.

---

Reviewers: 2 done | Synthesis: codex, 5s | Total: 10m13s

[roborev-ci]

roborev: Combined Review (db3867f)

Search refactor needs a pagination-stability fix before approval.

Medium

• internal/db/bun_search_content.go:890 — DuckDB leaves tool-call and result-event IDs null, causing tied rows to receive row_order = 0. Without further tie-breakers, paginated searches can duplicate, skip, or reorder results. Append call_index ASC, event_index ASC to the final ORDER BY.

---

Reviewers: 2 done | Synthesis: codex, 9s | Total: 12m8s

[roborev-ci]

roborev: Combined Review (e0be781)

One medium-severity ordering bug should be fixed before merging.

Medium

• internal/db/bun_search_content.go:820 — Content candidate ordering coalesces SQLite timestamps before converting legacy empty strings to NULL. An empty ended_at can therefore mask a valid started_at or created_at, causing recent substring or regex matches to sort last or appear on the wrong page. Normalize each timestamp before coalescing (for example, with NULLIF(..., '') in SQLite or a dialect-owned activity expression) and add an ordering test covering empty legacy timestamps.

---

Reviewers: 2 done | Synthesis: codex, 24s | Total: 21m23s