Spike: audit which contract rows and extension interfaces are ready to freeze at 1.0 #319

Context

Epic #56 defines 1.0.0 as the point where the extension interfaces lose
@experimental and the integration contract stops moving under a minor. Its
breakdown names the first piece of work: an audit that asks, for every contract
entry and every extension interface, whether it is ready to stop moving. The
audit's output becomes the epic's sub-issues. This is that audit.

What binds today, on main @ d350daa8:

• docs/integration-contract.md:3-14 — additive changes ship in a minor;
  breaking ones also ship in a minor before 1.0, with the breaking label, a
  maintainer ruling and a changeset, and need a major from 1.0. The reasoning is
  docs/decisions/contract-versioning.md.
• docs/integration-contract.md:573-576 — EmbeddingProvider,
  CompletionProvider, AgentRunner, SkillEmitter and the ui-sdk interfaces
  are @experimental until 1.0, with breaks announced as minor-version events.
• docs/extending/README.md:60-65 (Stability) and :69-106 (the seams): four
  in @schlessera/brain, four in @schlessera/brain-ui-sdk (AgentBackend,
  SpeechProvider, ToolRenderer, AsrClient), and SiteAdapter in
  @schlessera/brain-scrape (packages/scrape/src/adapter/types.ts:66).
• .changeset/config.json puts all fourteen packages in one fixed group,
  described in CONTRIBUTING.md:121-122 as a deliberate pre-1.0 tradeoff.

A 2026-09-22 comment on #56 found that the documented interface set is wider
than the @experimental annotations: AsrClient (packages/ui-sdk/src/client/asr.ts:15)
and SiteAdapter carry no individual tag. An audit that only greps for
@experimental misses them. git grep -l '@experimental' -- 'packages/*/src'
currently lists eight files.

Scope

• Inventory every surface 1.0 would freeze: each row of
  docs/integration-contract.md, each interface named in
  docs/extending/README.md, and each exported symbol tagged @experimental.
  Reconcile the three lists and name every surface that appears in one and not
  the others.
• For each surface, record: whether it has changed in a breaking way since
  0.33.0 (cite the CONTRACT: commit or changeset), whether a contract test
  exists that a third-party implementation could run, and a verdict: ready,
  needs work (what), or should leave the contract before 1.0.
• Say whether the fixed lockstep group is defended or reconsidered, with the
  argument, as the epic asks.

Out of scope

• Changing any interface, contract row, annotation or versioning policy. The
  audit produces a list; the list becomes Epic: 1.0.0 — the stability bar #56's sub-issues.
• Scheduling 1.0.
• New seams (ROADMAP binding decision 2).

Acceptance criteria

• A table posted on Epic: 1.0.0 — the stability bar #56 with one row per surface: where it is defined
  (file:line), where it is documented, its breaking-change history since
  0.33.0, whether a runnable contract test exists, and a verdict.
• Every mismatch between the three inventories is named.
• A stated position on the lockstep group, with its reasoning.
• Every "needs work" verdict is filed as a sub-issue of Epic: 1.0.0 — the stability bar #56, with template
  sections filled, or named in the table as not yet filed with the reason.

Verification

Another reader can re-derive the inventory from the three sources named in
Scope and reach the same row set.

Contract impact

• Neither box applies; the audit changes nothing.