Add a VALUE_ROUNDTRIP case-set to the driver-conformance census — "what you wrote is what you read back", enforced per driver per dialect (#12380 route D)

[os-warren]

Follow-up to #12380, ruled in by the maintainer 2026-08-26 alongside route A. Filed separately for sequencing reasons stated below.

⛔ Do not start until #12380's route-A fix has landed. This card would otherwise ship a gate that is red on SQLite by construction.

Why this exists — measured, not proposed

While investigating #12380, the dev was asked whether #12136's freshly-landed dialect axis had an opinion about a value-storage round-trip defect. The answer:

check-driver-conformance is green at 9 of 9 dialect-scored cells matrix-routed after #12136's MATRIXED invariant, but none of its nine case-sets is about value storage round-trip — so that green is not weak evidence about this defect, it is no evidence at all, and it would stay green forever with the defect in place.

The nine existing case-sets cover filters, temporal, pagination and aggregation. Nothing covers "what you wrote is what you read back." That is why this family keeps arriving one card at a time:

card	the same question, one instance at a time
#12380	SQLite Field.json is not round-trip faithful; PG/MySQL disagree
#11535	an array stored and read back as the string '["x","y"]'
#11065 / #11151	avg/sum over a boolean lowered to null/0, one driver over
#9276	a driver returning a declared field as an own key holding undefined

The route

Add a VALUE_ROUNDTRIP case-set to packages/spec/src/data/*-conformance.ts so the census's existing CLASSIFIED + MATRIXED invariants force every driver, on every dialect to answer the round-trip question. ⭐ This reuses machinery the census already has rather than building new — that is the main argument for it over a bespoke suite.

⚠️ Two constraints that shape this card

1. It is cross-driver by design, so it will ask driver-memory and driver-mongodb the same question. ⛔ [裁决] driver-memory / driver-mongodb 投入冻结 —— 维护者 2026-08-05 口径（跨单锚点） #5499 froze investment in both (maintainer ruling 2026-08-05, cross-card anchor). Read that ruling first and scope against it — the honest options are to ledger those drivers' cells with a reason, or to scope the case-set to drivers not under the freeze. Do not quietly start fixing frozen drivers, and do not let the freeze become an excuse to make the case-set vacuous.
2. A gate that ships red is worse than one that ships honest — this seat's own ruling on Resolve the 2 DIALECT_LEDGER rows, then promote "every dialect-scored cell has a matrix-routed suite" to an invariant (#12014 follow-up) #12136. Whatever this surfaces must be either fixed, ledgered with a named reason, or scoped out explicitly. ⚠️ And per Resolve the 2 DIALECT_LEDGER rows, then promote "every dialect-scored cell has a matrix-routed suite" to an invariant (#12014 follow-up) #12136, a DIALECT ledger row deliberately does not clear MATRIXED; check whether the same should hold for whatever ledger this needs, rather than re-opening an escape hatch that was just narrowed.

Sizing

⛔ Unknown, deliberately. #12380's route-A work will measure how many round-trip cells actually diverge on the SQL family; that number should size this card, exactly as #12014's number sized #12136 (which came out at 2). Re-measure at claim time rather than guessing here.

Clause ②

No — a census case-set plus its invariant. No shipped behaviour, no public surface. ⚠️ But note it will change what the census refuses, so any driver it newly reds is a real finding to route, not a number to tune away.

Refs

#12380 (parent; route A is the codec fix) · #12136 / #12014 (the dialect axis and MATRIXED this builds on) · #5499 (the driver-memory / driver-mongodb investment freeze) · #11535 · #11065 · #11151 · #9276 (the family this would close as a class)

[os-warren]

Claim: dev seat session session_01W6HFzyH98W1YaQXhJUJt6o, dispatched by the domain:engine PM seat, branch claude/issue-12393-value-roundtrip-caseset.

Blocker confirmed cleared before starting: #12380's route-A fix is on origin/main as 4045b954d3. Assignee untouched (PM already claimed).

Scope as briefed: a VALUE_ROUNDTRIP case-set in packages/spec/src/data/*-conformance.ts plus its census wiring. Sizing will be re-measured here, not inherited; the two named tripwires (#5499 driver freeze, and #12136's "a gate that ships red is worse than one that ships honest") are read and will be reported against explicitly.

---

Generated by Claude Code

[os-warren]

Dispatch stalled with zero work produced — un-claimed, back to pm:queue

Same stall as #12068 / #11671, but this one left nothing to recover. Measured at 09:00Z on the seat's own worktree: 0 commits ahead of origin/main, clean working tree, and origin/claude/issue-12393-value-roundtrip-caseset still at the base af56546adb. Nothing was written.

So the honest state is unclaimed, not "in progress": pm:dispatched + an assignee on a card with no work behind it is exactly the false signal that makes another agent skip it. Labels and assignee reverted to match reality.

The card itself is untouched and still fully specified — it is #12380 route D and stands as filed. It is not being re-dispatched in this batch only because the slot went to #12394 (autonumber high-water destruction, target:18), which is a live data-corruption card and outranks a coverage extension. Next free slot.

Third recurrence of the #12188 stall shape in this shift; the zero-output case is the sharpest instance of it so far, since nothing at all distinguishes a dead seat from a working one at the card level.

---

Generated by Claude Code