Fix: Ollama offerings resolve to the built-in OpenAI adapter, not the Ollama adapter

[TheGreatAxios]

Summary

CL-6586: sending a message to an Ollama-backed agent failed with
unrecoverable inference error: openai adapter: invalid quirks: default must be removed.

The catalog plugin for an Ollama connection is "openai-compatible" —
accurate, since @corbits/ollama-adapter wraps the built-in OpenAI
adapter unmodified. But resolveDefinitionSources sets
InferenceSource.provider to that same plugin value, and the
sidecar's adapter registry only recognizes a locally-served offering
under the key "ollama" (apps/sidecar/src/config.ts). A plugin-only
source therefore resolved to the built-in OpenAI adapter instead of
the registered Ollama one — whose stricter quirks schema has no
numCtx field and rejects the offering's quirks.default bag outright.

plugin (wire format) and the adapter-registry dispatch key are
different concepts that happened to collide for every provider but
Ollama. The fix keeps plugin as "openai-compatible" and corrects
only the in-flight InferenceSource.provider a launch pins into its
run config, in packages/folded-runs/src/launch.ts's deployAtHead —
the one and only place in the repo that consumes
resolveDefinitionSources's output. withOllamaAdapterKey rewrites
provider to "ollama" for any offering whose catalog provider is
actually named "ollama" (looked up via the stock, unmodified
listVisibleOfferings), leaving a caller-supplied sources override
untouched.

No vendored file changes, and ModelProviderPlugin stays a 4-value
enum — InferenceSource.provider is a free string, so this needed
nothing upstream.

Also corrects credential-test.ts's doc comment, which stated the
"plugin must be openai-compatible, never the provider's own id" rule
as absolute — true for every OpenAI-compatible provider but Ollama.

Verification

Live-verified against the real stack (.worktrees/cl-e2e-main,
tenant connected to a real Ollama instance over Tailscale):

• A pre-fix Myra instance still shows the original failure:
  This agent could not complete your request due to an unrecoverable inference error: openai adapter: invalid quirks: default must be removed.

• After applying this fix and creating a fresh "Just start talking"
  workbench (a new deployAtHead call, so it picks up the corrected
  code — an already-deployed run's bytes are baked at deploy time and
  don't pick up a code change), Myra's own opening greeting generated
  successfully, and a follow-up message got a real reply:

  Hi Myra, please confirm you're working by replying with one short
  sentence.

  Myra: Here and ready to help.

• Confirmed the outbound request actually carried options.num_ctx:
  querying the Ollama host directly (GET /api/ps) during the run
  showed gpt-oss:20b loaded with "context_length":131072 — the
  exact numCtx value from the offering's quirks table
  (OLLAMA_MODEL_DEFAULTS["gpt-oss:20b"]), not Ollama's own much
  smaller default context window.

Test plan

• bunx tsc --noEmit on packages/folded-runs and
  packages/hub-client (targeted, no repo-wide check)
• bun test packages/folded-runs/test/launch.test.ts — 27 pass,
  including a new test proving the provider-key correction and
  that a sources override is left untouched
• Live browser verification against a real Ollama instance (above)

Follow-ups

• packages/folded-runs is slated for deletion. Agent definitions,
  workflow, deployment, and versioning are moving native to Interchange.
  withOllamaAdapterKey lives in packages/folded-runs/src/launch.ts
  because deployAtHead is the one call site today — that placement is
  right for now, but it must move with that cutover, not get dropped
  on the way, or this bug comes back silently.
• CL-6588:
  an agent deployed before this fix never picks it up — a launch
  renders its bytes once, and nothing re-renders them when the fix
  lands. Proven live: a pre-fix Myra instance in the same tenant still
  errors after this fix shipped, while a freshly created one works.
  Out of scope here; tracked separately.