feat(evaluation): conversational correction challenge corpus (CC-01..CC-15)

[obvious-autobuild]

Acceptance criteria

1. 15 independent conversational-correction cases (CC-01..CC-15) as data-only fixtures — six case-group files plus a shared synthetic household env (3 children, 4 members, explicit read/read-write grants, America/New_York timezone). Families: time corrections, wrong child, wrong prior event, multiple authors / stale suppression, negation (including zero-care capture), repeated corrections / undo.
2. Candidate-agnostic adapter contract with an environment-dependent factory (createCorrectionAdapter(env)); a deterministic runner drives every candidate identically and resolves each turn to applied / rejected / clarification.
3. Reference adapter passes 15/15 (mechanical rule interpreter — regex dispatch, lexical child/entry resolution, snapshot-stack undo, timezone-safe retargeting via Intl; no I/O, no LLM, no clock).
4. Negative control fails the corpus: the reference with five named faults (duplicate-on-retarget, lineage truncation, stale clobber, skipped grant check, suppressed zero-care write) is flagged on 7/15 cases; --expect-failure inverts the exit code.
5. Mechanical enforcement per turn and at final state: raw transcript immutability, original-author stability, append-only revision lineage (kind + author per entry), entry-count deltas, publication as the only status transition and runner-driven, replayed correctionId = applied no-op, fail-closed write grants, undo = restore state before last mutation with an appended undo revision, zero-care = captured reviewable record (never a 0-duration quantity), ambiguous referents clarify with zero writes.
6. packages/domain untouched — proposal-level lineage shapes; scope documented in the README section and package scripts (eval:corrections, eval:corrections:broken).

Verification evidence (exact head under review)

Head SHA: c32e79a — all results below ran against this commit in a clean worktree off origin/master (652ace5):

Check	Result
bun x tsc --noEmit	exit 0 (strict, noUncheckedIndexedAccess, verbatimModuleSyntax)
bun src/corrections/run.ts	15/15 cases PASS, exit 0
bun src/corrections/run.ts --adapter=./src/corrections/example/broken-adapter.ts	7/15 (one flag per fault), exit 1
same + --expect-failure	faults caught, exit 0
bun src/run.ts (capture corpus regression)	6/6 fixtures PASS, exit 0

Scope

• Adds only evaluation/corrections/**, evaluation/src/corrections/**, and edits evaluation/README.md + evaluation/package.json (two scripts + README section).
• evaluation/ is not a pnpm workspace member, so no lockfile change and no product/backend surface touched.
• Companion to the merged capture corpus (PR feat(evaluation): shared acceptance corpus + candidate-agnostic cross-review harness #6); the correction problem's product implementation stays out of scope until the contract lane lands lineage schemas.

[obvious-autobuild]

Evidence receipt

PR:               https://github.com/OCPdev25/obv-hackaton/pull/19
Tested head SHA:  8d386deb37a8cd4dc7a9b59a0aa3e7201dc97ca4
                  (rebased head — original c32e79a invalidated by the PR #22/#24
                  base movement per workflow rule 5; full battery re-run at 8d386de)
Review result:    pass — author review per repo policy (no required human
                  approvals); full diff read against the six acceptance criteria
                  in the PR body
Checks:           all green — "Typecheck, test, build" pass on 8d386de
                  (run 35260948598, 42s); PR CLEAN/MERGEABLE pre-merge
Merge commit:     69e7bd3 (squash onto master)
Post-merge smoke: on 69e7bd3 — pnpm install --frozen-lockfile ok;
                  pnpm typecheck ok; pnpm test ok; pnpm build ok;
                  bun test ./security 17 pass / 0 fail;
                  bun src/run.ts 6/6; capture broken control --expect-failure ok;
                  bun src/corrections/run.ts 15/15 (reference-rule-adapter);
                  corrections broken control 7/15 flagged, --expect-failure ok;
                  bun src/manifest-selfcheck.ts ok (no fixture-area interference)
Unlocked tasks:   slot 24 (append-only corrections / stale-result) can consume
                  this corpus as its candidate-agnostic gate; cross-review can
                  now score spoken-correction adapters on both corpora

Conflict resolution note (rule 5 rebase onto 44856c6): the two conflicts with
PR #22 were resolved as unions — package.json keeps both selfcheck:manifest
and the eval:corrections* scripts; the README keeps the manifest docs plus the
corrections section. The corrections corpus is deliberately not a
manifest-registered area (own runner + turn-protocol adapter contract); the
future correction-protocol fixture class is the documented seam if one CLI is
wanted later.