Finish adapter correctness and release verification for the v2 removal

[tobyhede]

Part of #778 and the EQL v2 removal. This combines independent adapter correctness fixes with the release gates that prove the removed surface has no surviving consumer.

Acceptance criteria:

• Preserve nested PostgREST OR structure and referencedTable for encrypted string and structured or filters; no plaintext operand may reach the wire. Reject or deliberately handle nullish matches() operands instead of degrading them to plaintext filters.
• Define safe behavior for non-row result modes such as .csv() so CSV text is never passed to bulkDecryptModels.
• Honor SQL LIKE escaping when adapting encrypted like / ilike calls: escaped % and _ are literals, not wildcard controls.
• Detect the removed legacy encryptedSupabase({ encryptionClient, supabaseClient }) call shape (or otherwise validate the supplied client) and fail with an actionable migration error before the first query.
• Align native, WASM, and Supabase Date reconstruction, including invalid values, nested paths, aliases, and bulk behavior. Relates to decrypt / bulkDecrypt hand back date columns as ISO strings while the model helpers return Date — silently, and not for the stated reason #779; do not widen decrypt or bulkDecrypt plaintext input types here.
• Typecheck examples, e2e consumers, and stack-drizzle integration sources in CI. The e2e package must declare the repository-pinned TypeScript dependency rather than resolving an ambient/global tsc.
• Correct the stash init Supabase handoff to name the concrete public.eql_v3_* domain family (with a valid example), and update its unit test and changeset together; eql_v3_encrypted does not exist.
• Reconcile pending changesets, skills, READMEs, and the feat!: remove EQL v2 repo-wide (umbrella, #707) #772/EQL v2 removal: review remediation for #772 #778 tracking text after the removal work lands. Do not rewrite already-published historical changesets. The audit must specifically ensure:
  • migration guidance names only the final 1.0 imports and symbols (no removed /v3 subpaths or *V3 adapter names, and no superseded Supabase behavior);
  • adapter behavior changes target the final owning adapter package in changeset frontmatter, not @cipherstash/stack by default;
  • the DynamoDB entry reflects decrypt audit support on the typed Encryption client; and
  • @cipherstash/nextjs receives only package-specific release notes, not the unrelated deleted-package narrative.
• Make CLI and wizard integration skill selection share one source of truth, or add a parity test that covers every integration and verifies SKILL.md availability, including Prisma Next.

The adapter fixes may be separate commits, but share one release-readiness outcome and final verification pass.

[tobyhede]

Status (2026-07-30): 8 of 9 criteria delivered — criterion 8 blocks closure

Completed by #828, merged as ee33ac3b, an ancestor of origin/main (c8b1325a), reaching main inside the #772 umbrella merge.

Note the tracking failure: #828's body ends with Closes cipherstash/stack#816, but it targeted the remove-v2 branch. GitHub only honours closing keywords on merge into the default branch, so it never fired and closingIssuesReferences is empty. Same mechanism left #814 stale-open (now closed).

Verified delivered against origin/main:

• Nested PostgREST OR + referencedTable preserved, no plaintext operand on the wire — packages/stack-supabase/src/query-filters.ts:334-360
• Nullish matches() rejected — query-builder.ts:317-321
• .csv() made safe — query-builder.ts:496-499, types.ts:602-603
• LIKE escaping honoured — like-pattern.ts:18-46, rejection at query-builder.ts:871-877
• Legacy encryptedSupabase({ encryptionClient, supabaseClient }) shape diagnosed — stack-supabase/src/index.ts:149-152
• Native / WASM / Supabase Date parity — packages/stack/src/eql/v3/date-reconstruction.ts:12-22, 26-70, consumed at client-v3.ts:298, wasm-inline.ts:1412, dynamodb/operations/decrypt-model.ts:81, bulk-decrypt-models.ts:85, stack-supabase/src/query-results.ts:80-83
• Examples / e2e / stack-drizzle integration typechecks gated, e2e pins TypeScript — .github/workflows/tests.yml:161,184,187 and :124-128; e2e/package.json devDep typescript: catalog:repo
• stash init Supabase handoff names the concrete domain — init/lib/setup-prompt.ts:103, asserted at setup-prompt.test.ts:196-197 including not.toContain('eql_v3_encrypted')
• CLI/wizard skill parity including Prisma Next — e2e/tests/skill-map-parity.e2e.test.ts:37 (the criterion allowed a parity test as an alternative to one source of truth)

Outstanding: criterion 8, the README half

Reconcile pending changesets, skills, READMEs, and the #772/#778 tracking text after the removal work lands.

The changeset and tracking-text work is done. Two published READMEs still describe removed tooling:

• packages/stack/README.md:576 — table row | --proxy/--no-proxy | Opt in/out of the CipherStash Proxy path |. The flags were removed in feat(cli): remove the EQL v2 migration leaf #825; packages/cli/README.md, skills/stash-cli/SKILL.md and registry.ts are all clean, and v2-retirement.test.ts:11,31-32 asserts their absence. This one line was missed.
• packages/wizard/README.md:22-24 and :35-36 — still say the wizard "shells out to stash eql install / db push". Only the db push half is wrong; stash eql install is still live (registry.ts:307). packages/wizard/src/lib/post-agent.ts:58 now states the retired command "must never run", so the README contradicts the shipped code.

Do not close this by merging #790. #790 does remove the --proxy row, but it is CONFLICTING, based 158 commits behind main, and its two-dot diff against main is 389 files / ~19,300 deletions — merging it unrebased would revert large amounts of already-merged work. The one line needs doing directly.

Wizard source and test db push residue is out of scope here (it is not a README) and is tracked in #837.