feat: declare per-model capabilities on session/new

[zxch3n]

Why

configOptions is rebuilt for the current model: fast-mode appears only while
that model has a fast speed tier, and reasoning_effort carries that model's
list. A client reading the session/new snapshot therefore learns nothing about
any other model — and ACP has no request that asks.

Lody was reading those options as an agent-wide capability catalog, so selecting
a model other than the probed one produced answers like "this agent has no fast
mode" for a model that does, or the reverse.

What

Attach the catalog this server already holds to the session/new response:

{
  "configOptions": [ /* unchanged */ ],
  "_meta": {
    "lody": {
      "modelCapabilities": {
        "version": 1,
        "models": {
          "gpt-5.2":      { "effortValues": ["low", "medium"], "fastMode": false },
          "gpt-5.6-luna": { "effortValues": ["low", "medium", "high", "xhigh"], "fastMode": true }
        }
      }
    }
  }
}

Built from the same sessionState.availableModels the config options are built
from — supportedReasoningEfforts and modelSupportsFast(model). No new source
of truth, no extra request, nothing fetched.

Contract

Self-declared and advisory. It may describe a control the snapshot does not
carry; it never grants permission, authorizes a value, or rejects one. The live
session state remains the authority.

A model absent from models is unknown, not unsupported. The consumer
therefore reads the catalog whole or ignores it whole, and treats it as stale
after a TTL or a change of adapter version.

Emitted only when session config is enabled and the model list is non-empty, so
a client that ignores _meta sees a byte-identical response.

Consumer

Read by Lody in LodyAI/Lody#333, which ships first and reads an absent
declaration as "nothing to say" — so this can merge on its own schedule.

Test

npx tsc --noEmit clean.

🤖 Generated with Claude Code