Migrate conditional-lookup onto a new field-add lifecycle driver

[HynLcc]

What

Migrates the conditional-lookup runner kind (lookup/conditional-10k) onto a brand-new shared driver, field-add-lifecycle.ts.

conditional-lookup fits neither existing driver:

• No single prepare phase — its create/seed measurements ride the fixture with cache-variant names (seedBuild/createTables/seedRestore), like field-convert-lifecycle but unlike record-mutation-lifecycle.
• Two primary phases (createLookupField + fullLookupScanReady) feeding a computed conditionalLookupReadyMs threshold — unlike field-convert-lifecycle's single metric-named phase.
• Restore-or-delete cleanup (drop the added lookup field) — unlike field-convert's keep-or-delete (in-place, non-invertible).

So this introduces the sibling family driver for runners whose measured op adds a field over a seeded table, waits for the new column to backfill, then restores the seed by deleting that field. field-duplicate and field-create ride it next (rounds 2–3), generalizing it as the family migrates.

The driver owns the repeated protocol (the seedReady phase, diagnostic-error wrapping, cleanup invocation). conditional-lookup declares only what varies: the dual-table seed fixture, the seed assertion, the trace-wrapped field-add + full-scan readiness bundle, the result assembly (reusing buildConditionalLookupCaseResult unchanged), and the restore-or-delete cleanup. The exported seed helpers field-duplicate reuses are untouched.

diff-artifacts.mjs masks

• conditional-lookup's generated ids/names (source/host table ids, sourceFields/hostFields ids, host record ids, table names) — proven volatile by an unmodified-main baseline A vs B (the case was never diff-validated before).
• details.seed.seedHash/seedHashShort — the same content-addressed seed-cache key every other migrated runner already masks under a cache object. It digests the seed code files, so it moves on a pure refactor while staying identical run-to-run; conditional-lookup just placed it directly under details.seed.

Validation (local, lookup/conditional-10k, v1+v2)

• Baseline A vs B clean after the noise masks.
• pnpm check green.
• G1 baseline vs candidate clean for both engines (only the documented seedHash key masked).
• Negative tests: phase / verifiedSamples.expected / threshold metric / recordCount / seedNamePrefix drift all still caught; only the seedHash key is ignored.
• Trace ref count unchanged (23 both engines).
• Semantics: both engines pass, 6 phases match baseline, conditionalLookupReadyMs = createLookupField + fullScan, full 10k scan verified, no routing assertion (correct for a computed-lookup case).
• New driver — no other members to regress this round.

Tracker: 20/35 runner kinds, 24/55 cases.

🤖 Generated with Claude Code