Migrate formula-table, record-update-attachment, duplicate-table onto lifecycle drivers

[HynLcc]

What

Migrate three legacy runner kinds onto lifecycle drivers, taking the tracker
from 23/35 → 26/35 runner kinds (31/55 → 36/55 cases). Each migrated runner
delegates the repeated protocol to a driver and reuses its UNCHANGED
result-assembly, proven byte-equivalent by the G1 artifact diff over every
case × engine.

Runner	Driver decision	Why
formula-table	Reuse field-add-lifecycle.ts (4th member)	Same family as field-create: seed a table → add field(s) → wait computed backfill → verify → restore-by-delete. The sourceReady-vs-seedReady phase relabel and trailing formulaReady phase append are expressed in the spec; driver unchanged.
record-update-attachment	Reuse record-mutation-lifecycle.ts (useRecordWindow:false)	Same family as record-update: bulk update over seeded rows. The execute-only token upload + warmup + p95 sampling live in the measured op (NOT prepareFixture — tokens must never be cached: seed→execute carries no storage volume); driver unchanged.
duplicate-table	New duplicate-lifecycle.ts (1st member)	No existing driver fits "duplicate an entire seeded entity": field-add's cleanup is restore-by-delete, but duplicate must DROP the created copy. New minimal, family-shaped driver (duplicate-base is the deferred 2nd member).

No shared driver was modified — field-add-lifecycle.ts and
record-mutation-lifecycle.ts stay byte-stable, so their existing members
(conditional-lookup, field-duplicate, field-create / record-update,
record-create, record-reorder, selection-clear, record-update-link) need no
re-validation.

G1 evidence (baseline A on unmodified main ↔ candidate, local v1+v2)

All 10 case × engine artifacts diff CLEAN. result=pass, error=null on every one.

Case	Engine	result	route (v1=false/v2=true)	primary metric	trace refs (cand==base)
formula/10k-calc	v1	pass	n/a (no routing assert)	formulaFullReadyMs 1238ms	12==12
formula/10k-calc	v2	pass	n/a	formulaFullReadyMs 1128ms	12==12
formula/10k-5-concurrent	v1	pass	n/a	formulasFullReadyMs 3422ms	16==16
formula/10k-5-concurrent	v2	pass	n/a	formulasFullReadyMs 1919ms	16==16
record-update/attachment-insert-100	v1	pass	routeMatched, v2=false	p95 95ms	21==21
record-update/attachment-insert-100	v2	pass	routeMatched, v2=true	p95 240ms	21==21
duplicate-table/10k-20f	v1	pass	routeMatched, v2=false	duplicateTableRequestMs 335ms	12==12
duplicate-table/10k-20f	v2	pass	routeMatched, v2=true	duplicateTableRequestMs 7870ms	12==12
duplicate-table/10k-25f-5formula	v1	pass	routeMatched, v2=false	duplicateTableRequestMs 599ms	12==12
duplicate-table/10k-25f-5formula	v2	pass	routeMatched, v2=true	duplicateTableRequestMs 9828ms	12==12

Baseline A↔B noise check is clean with the existing + added noise masks; the
seedHash-family difference appears ONLY in G1 (never in A↔B), as expected.

Mask deltas (scripts/diff-artifacts.mjs)

• formula-table — seedHash family: extended the details.seed rule to mask
  seedTableName (its suffix IS the hash; already masked seedHash/seedHashShort).
  Noise: details.{formula,formulas[],formulaResults[]}.compiledExpression
  (embeds generated A/B/C field ids).
• record-update-attachment — no new seedHash mask (its hash nests under the
  already-masked details.seed.cache). Noise: details.request.attachmentFieldId,
  details.update.expectedTokens.
• duplicate-table — no new seedHash mask (hash nests under the already-masked
  details.prepare.cache). Noise: details.sourceFields[].id,
  details.sourceFormulas[].id, details.duplicate.duplicatedFormulaFields[].id.

Every noise mask is justified by the baseline A↔B diff (volatile on UNCHANGED
code); field NAMES, expected values, counts, routing, and verification evidence
stay visible.

Negative tests (diff has teeth) — all 9 behaved correctly

• (a) semantic perturbation → FAIL: formula verifiedSamples[].expected,
  duplicate rowCount, attachment rowCount each made the diff FAIL.
• (b) masked-field change → PASS: formula seed.seedTableName, duplicate
  sourceFields[].id, attachment update.expectedTokens each kept the diff PASS.
• (c) over-mask guard (unmasked sibling) → FAIL: formula seed.seedNamePrefix
  (sibling of masked seedHash/seedTableName), duplicate sourceFields[].name
  (sibling of masked id), attachment update.updatedRecords (sibling of masked
  expectedTokens) each still made the diff FAIL.

Verification

• pnpm check passes.
• Trace refs preserved (candidate traceRefCount == baseline on every case;
  savedTraceCount=0 locally — no Jaeger — CI is where saved==ref holds).

🤖 Generated with Claude Code