Skip to content
Source
Sign inSign up
TanStack/db

Oracle: revalidate retained applied-receipt and lazy child-ready receiving gaps #2342

Send to my agent
OpenKyleAMathews(GitHub profile, opens in a new tab) opened this issue 1 comment

Retained obligations

The release deferral record retains two receiving gaps after the P06/P07/P13/P14 prose corrections:

  • Release P08: exercise the actual custom-adapter fenced callback while it holds an establishing applied receipt before load settlement.
  • Release P14: compare the actual lazy child-ready call outcome with the existing source-health model, rather than catching and discarding a throw.

These are historical coverage obligations, not newly reproduced production defects. Revalidate each against current main before implementing anything; later work may already discharge one. The release P08 identifier here is not the separate metadata P8 / #2317 investigation.

Per-cut acceptance

  1. Identify the current primary executable owner through the coverage map and read its contract and limits.
  2. Preserve the accepted owner, fresh-run neighbors and omitted-callback reach controls from the retained campaign.
  3. Add the actual receiving observation to that owner's model/driver/comparison, with legal generated histories and measured reach.
  4. Demonstrate sensitivity by omitting or corrupting the actual callback/outcome and showing that the intended comparison rejects it. Catching an error without judging its outcome supplies no receiving evidence.
  5. Preserve current laws and budgets. If a production failure is discovered, show parent RED before repair, repaired GREEN and a hostile mutant, and state the bounded class covered.

Record each cut's disposition independently. A source/prose correction alone does not discharge an executable receiving obligation.

Related: #2118 covers broader shared adapter settlement/readiness conformance. This issue keeps these two concrete retained release cuts visible without treating that larger proposal as completed evidence.

  1. Fresh revalidation on freshly fetched origin/main baabf57a0f110d57ad9bb7b81148466b18fbe0d1: both retained cuts remain open.

    Release P08 — actual documented/custom guidance callback: the current packages/db/skills/db-core/custom-adapter/SKILL.md on-demand fence calls commit() without awaiting its receipt. A fresh probe extracts and executes that exact fenced callback through a real Collection, invoked outside the independent mutation holding application. At the cut: one actual commit, receipt pending, row invisible, but load already fulfilled. The applied-settlement comparison fails (fulfilled versus required pending). The same extracted callback with only an awaited receipt passes; an omitted-callback observation is rejected before granting reach. This is a demonstrated guidance callback violation of LoadSubsetFn/architecture law 7, not a claim about a shipped Query or persistence adapter. No guidance or production repair was made. The fixed diagnostic does not supply the missing committed generated receiver. This is unrelated to metadata P8/#2317.

    Release P14 — lazy child-ready outcome: the unchanged lazy owner passes 71 tests, but still catches/discards the actual call's exception. Four additional scratch histories reuse its shared model/driver and preserve failed-first-load, fresh-run, reload and absent-callback neighbors. An in-memory overlay throws only after actual present callbacks execute; the existing comparison still passes all 75 tests, including two actual failed-first-load callback throws. All 21 existing lazy reach states remain reached. This is a surviving wrong-outcome mutant demonstrating the observation gap, not a baseline production RED. Absent optional callbacks receive no actual-call reach credit; generated fail→actual-ready coverage still needs to be added.

    The existing receipt-refinement and documented-mutation neighbor controls also pass (15 tests). Keep both cuts open until the actual receiving observations are integrated into their owners, with generated per-cut reach and sensitivity that rejects the wrong callback/outcome. Production and tracked tests stayed unchanged; no broad adapter or #2118 conformance closure is claimed.

Sign in to join this conversation.