ci: add npm publish workflow via OIDC trusted publishing

Release 0.5.4

Release 0.5.3

Release 0.5.2

Release 0.5.1

Release 0.5.0

chore(0.4.2): make persistent memory actually persist

The MemoryExtractionMiddleware scheduled fact extraction as a
fire-and-forget asyncio task, but _execute's finally block closed the
shared httpx client first — producing the warning 'Memory extraction
failed: Cannot send a request, as the client has been closed' on every
`prax prompt` run and writing nothing to .prax/memory.json.

_execute now drains the pending extraction task (bounded at 15s) before
closing the client. Additionally, the extraction call was using the
non-streaming `complete()` path which some OpenAI-compatible proxies
reject with `400 Stream must be set to true`; extraction now uses the
streaming transport when the provider declares supports_streaming, via
a new shared helper _extraction_llm_call().

Verified end-to-end: seed 4 facts across categories (preference,
behavior, goal), then three independent `prax prompt` sessions correctly
recall the name / language / project / off-hours boundary — 3/3 hit
rate, zero memory warnings, and the store even grows incrementally as
later sessions surface new preferences.

chore(0.4.1): fix provider lookup collision that broke README Quick S…

…tart

When a user's .prax/models.yaml reuses a bundled model name (e.g. a proxy
provider configured with model `gpt-5.4` or `glm-4-flash`), the first-scanned
bundled provider won the lookup even though its credential env var was unset.
Result: a confusing 'RuntimeError: No configured models are currently
available' in the catalog path and a cryptic 'LocalProtocolError: Illegal
header value b'Bearer '' in the outbound HTTP path.

Both resolution sites (core/model_catalog.get_model_entry and
core/llm_client.resolve_model) now prefer the first provider whose api_key
env resolves to a real value, falling back to the first match only when no
credentialed candidate exists. Unit tests lock in the new behavior.

Also demoted the 'Docker sandbox unavailable — falling back to local sandbox'
log from WARNING to INFO so first-time users without Docker don't see a red
warning on every run. fail_closed policy still errors loudly.

docs(getting-started): correct verification step to use 'prax providers'

The previous step told users to run 'prax doctor' and showed made-up
output like '[glm] ✓ ZHIPU_API_KEY is set'. That output is fabricated:
'prax doctor' is the Claude-Code integration doctor, not the provider
health check. Swap in 'prax providers', which is the actual CLI entry
and matches the real output format ('<model> (<tier>, available, ...)').

chore: bump version to 0.3.2