fix(code-mode): attach learner_id to compat-proxy calls and surface upstream 402 as the credit UX

[null-crafter]

Checklist

• Tests were added/updated according to the feature/bugfix/change made
• Version was rolled according to semver requirements
• API endpoints openapi schema was updated if applicable

Changes

What

Two related fixes to the Code-mode (opencode) OpenAI-compat loopback proxy, so
desktop model calls are attributed to the signed-in learner and respect credit
exhaustion the same way normal chat does.

1. learner_id query param on every forwarded model call

The proxy now appends learner_id=<username> to each request it forwards to
…/api/ai-mentor/orgs/<tenant>/v1/*, so upstream usage lands on the right
learner:

• The username reaches Rust via a new set_opencode_learner command, called
  from an app-root hook (hooks/use-opencode-learner.ts, mounted in
  providers/index.tsx) — reactive to login/logout via useUsername(); an
  empty username clears a stale learner. App-global by design: a desktop
  install has exactly one signed-in user.
• Read per request (not captured at session spawn), so a login that lands
  while a session is alive takes effect on its next call.
• The agent controls its own query string, so any agent-supplied learner_id
  is stripped before the real one is appended (names are form-encoded — a
  hostile username can't smuggle extra params). The proxy stays a credential
  boundary: attribution comes from the app, never from the agent.
• New command registered in both entrypoints (main.rs + lib.rs) and added
  to the ACL (permissions/default.toml block + default set,
  capabilities/default.json).

2. Insufficient credit (402) forwarded and shown as the normal-chat credit UX

The backend is starting to answer 402 on chat/completions at credit 0.
Without this change the agent reduces that to an opaque error string (generic
support toast) or swallows it entirely (silent empty turn).

• The proxy still relays the 402 + JSON body to the agent verbatim (new
  end-to-end regression test proves status, body, and the auth swap), and now
  also emits opencode:payment-required to the webview with the parsed body.
• A new app-root hook (hooks/use-opencode-402.ts) listens and drives the
  existing handle402Error — insufficient-balance toast, then the Stripe
  UpgradePackageModal / admin billing redirect / contact-admin banner,
  exactly as the WebSocket chat path does. If the 402 body lacks the Stripe
  pricing table, it's backfilled cache-first from the account-billing query
  (CreditBalance keeps that cache warm).
• Retry bursts collapse into one toast+dialog, and wasRecent402() suppresses
  the generic chat error toast for the same failure — mirroring normal chat,
  which returns before its errorHandler on 402.
• Debug builds also log each forwarded request and its response status to the
  tauri:dev terminal (payload capped at 4 KiB; never in release — the
  payload is the user's chat).

Also: coding-mode-button.tsx now caches one shared dynamic import of
@tauri-apps/api/core — two concurrent first-time imports could race under
vitest's module aliasing and hand the loser the unmocked module.

Testing

• Rust: cargo test opencode_proxy — 10 tests, including URL-building
  (append/strip/encode learner_id) and a live end-to-end 402 passthrough
  against a canned upstream.
• Vitest: new suites for both hooks (learner sync 5, 402 UX 8 — backfill,
  burst dedupe, suppression window, non-Tauri no-op); full unit suite green
  (405 files / 11k+ tests); coverage on changed files ≥95 % lines
  (hooks at 100 % / 96.8 %).
• Typecheck, ESLint, Prettier, rustfmt, and both e2e journey-coverage checks
  Note for manual verification: pnpm tauri:dev loads the deployed frontend
  (devUrl), so the frontend half needs
  cargo tauri dev --config '{"build":{"devUrl":"http://127.0.0.1:3000"}}'
  against a local pnpm dev, or a deploy. The Rust half is visible immediately
  (learner set: <user> at load, ← 402 POST … on a credit failure).

🤖 Generated with Claude Code

[github-actions]

PR E2E - timed_out

result	timed_out
central run	31819862077
commit	deb84d18

No test report was produced (result: timed_out). The run did not finish a suite — usually because another PR's run took the shared environment, a newer commit/label superseded this one, or it was abandoned before completing. This says nothing about your code.

Re-trigger it yourself: comment /retest (or /retest <spec> to re-run a single file in minutes). /status shows the queue. (Avoid Re-run jobs — that replays the old merge ref instead of recomputing it against current main.)

Last completed run for this PR: report — still valid unless this PR has changed since.

[github-actions]

PR E2E - FAILED

Full report, traces and screenshots

result	FAILED
central run	31529481313
commit	8a232696
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	2
new failures	2

[github-actions]

PR E2E - FAILED

Full report, traces and screenshots

result	FAILED
central run	31602594289
commit	10047a83
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	3
new failures	3

[github-actions]

PR E2E - FAILED

Full report, traces and screenshots

result	FAILED
central run	31612998291
commit	10047a83
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	3
new failures	3

[github-actions]

PR E2E - FAILED

Full report, traces and screenshots

result	FAILED
central run	31711765603
commit	deb84d18
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	1
new failures	1

[github-actions]

PR E2E - FAILED

Full report, traces and screenshots

result	FAILED
central run	32058342041
commit	deb84d18
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	4
new failures	3

[github-actions]

PR E2E - OK

Full report, traces and screenshots

result	OK
central run	32145804870
commit	51c4104e
suite	mentor - mode full - env stg1
tested against	prod release 1.126.0 (env was synced to it first)
failed	0
new failures	0