fix: discovery improvements, dead-code removal, tests, and docs alignment (#24)

* fix: improve discovery correctness and scope model cache per provider

- Format dash-separated date stamps correctly: `claude-opus-4-5-20251101`
  now renders "Claude Opus 4.5 20251101" instead of "Claude Opus 4 5 20251101"
- Fall back to `max_tokens` for the context limit when `max_input_tokens`
  is absent, so deployments reporting only the OpenAI-style total get limits
- Map cache read/write per-token costs from `/v1/model/info` into
  OpenCode's `cacheRead`/`cacheWrite` cost fields
- Key the SWR model cache by `providerId@baseURL` so two providers
  pointing at the same proxy (possibly with different API keys) no
  longer share one model list

* refactor!: remove unused transport-routing code and @opencode-ai/sdk dependency

The V2 transport-bucketing path (`discoverBucket`, `buildModelV2`) has
been unreachable since 0.5.0 switched fully to the config hook; nothing
imported it. Delete it along with its orphaned helpers and types, and
drop the `@opencode-ai/sdk` dependency it was the only consumer of —
the "zero runtime deps beyond @opencode-ai/plugin" claim is now true.
Also adds the vitest dev dependency and `test` script for the suite
added in a follow-up commit.

BREAKING CHANGE: remove previously exported but dead public types
(`Transport`, `TransportPolicy`, `LiteLLMOptions`, `ModelType`). No
behavior change — the config-hook plugin is untouched.

* test: add vitest suite and run it in CI

Cover the pure logic that regresses silently: model-name formatting
(including version-pair/date-stamp edge cases), model categorization,
base-URL normalization, and the SWR disk cache (round-trip, per-key
isolation, version mismatch, max-age expiry). Tests live in `test/`
so they stay out of the published tarball. CI now runs `npm test`
alongside typecheck on Node 20/22.

* docs: align README and CONTRIBUTING with current behavior

The README still documented the dual-provider architecture removed in
0.5.0: the `litellm-responses` sibling provider, `transport` /
`responsesApiModels` / `chatApiModels` options, transport bucketing in
the How-it-works diagram, and `organizationOwner` extraction — none of
which the config-hook plugin does. The FAQ even pointed users at
`responsesApiModels` to fix a reasoning_effort error, which silently
did nothing.

Rewrite those sections around the actual single-provider + SWR-cache
flow, fix the naming example ("Claude 3.5 Sonnet"), document the cache
and background refresh, trim shipped items from the roadmap, update the
source tree, drop the redundant quickstart install step, and fix the
plugin log path in CONTRIBUTING.

* fix: use OpenCode's snake_case cache cost field names

The config schema at opencode.ai/config.json defines the model cost
object as { input, output, cache_read, cache_write }; the camelCase
cacheRead/cacheWrite keys emitted previously would be silently ignored.

* docs: correct offline-proxy FAQ about warm-cache behavior

A warm on-disk cache is served without any network call, so discovered
models keep working when the proxy is down; the old answer described
pre-cache behavior.

* test: restore original XDG_CACHE_HOME after cache tests

* docs: add architecture banner to README

Hand-crafted SVG showing the config-hook flow: OpenCode -> plugin
(discover/merge/cache) -> LiteLLM proxy -> upstream providers, plus the
SWR disk cache. Dark, crisp at any size, and no diagram tooling to
maintain.

* fix: scope in-memory discovery state by provider cache key

Key refreshContexts, refreshInFlight, and injectedModelIds by
`providerId@baseURL` instead of the bare baseURL, so two providers
pointing at the same proxy (with different keys) can no longer
suppress each other's model injection or reuse the wrong refresh
context.

* fix: refuse version-pair merges inside numeric runs

`model-1-2-3` previously rendered as "Model 1 2.3" because only the
token after the pair was checked. Also refuse the merge when the
token immediately before it is a short number, keeping numeric runs
like `1-2-3` unmerged.

* docs: correct health-check placement in sequence diagram

With a configured baseURL the proxy is not contacted during startup
unless the cache is cold — the only 3 s fail-fast health check is the
port probe during auto-detection. Move the contact point below the
cache read so the diagram matches the SWR fast path.