fix(dash): chat /v1 404 + Esc closes a running serve modal

[michaelroy-amd]

Summary

Two dash chat usability bugs hit when serving a model and then chatting:

1. Chat completions 404 (completion error, http error, 404 detail not found).
   The Rig OpenAI client builds the request URL as base_url + "/" + "chat/completions",
   so base_url must end in /v1. Every endpoint carried it (VLLM_ENDPOINT,
   LEMONADE_ENDPOINT, the managed-service endpoint_url) except the probed
   default DEFAULT_CHAT_BASE_URL = "http://127.0.0.1:8000". So when the dash
   probed a vLLM already listening on :8000 and fell back to that default, it POSTed
   to /chat/completions and the server answered 404 {"detail":"Not Found"}.
   Fix: add the /v1 suffix so it matches the other endpoints.

2. Serve modal felt stuck. A managed serve blocks for up to 45s in its HTTP
   readiness wait (wait_for_service_http_ready). During that window the job console
   is Running, where Esc/Enter were swallowed and only q closed — but the
   footer advertised only Ctrl+C cancel, so there was no visible exit and it
   looked frozen. Fix: Esc now closes the overlay while a job is running (the
   job keeps running in the background, same as q already did), and the footer
   says so. Finished-job Esc/Enter still dismiss the console back to the screen
   body; Enter stays inert while running to avoid accidental dismissal.

Scope of the Esc-while-running change

The change lives in the shared job_console::on_console_key, so it applies to
every overlay that hosts a job console — 12 callers in total: serve, examine,
engine, runtime, install, update, config, logs, command, automations, services,
and onboarding.

Behavior change worth flagging: Esc on a running onboarding step
(onboarding.rs:180, *ob = None) now closes the whole onboarding wizard. This
is consistent with what q already did there, and it's a deliberate decision —
a running step keeps running in the background. Every caller's Unhandled arm was
checked; none branch on Esc, so nothing screen-specific is silently intercepted.

Changes

• llm.rs: DEFAULT_CHAT_BASE_URL → http://127.0.0.1:8000/v1 (+ rationale doc),
  plus a revert-proof regression test asserting the resolved base_url ends in /v1.
• ui/job_console.rs: Esc on a running job → Closed; updated footer hint.
• ui/serve_wizard.rs, ui/examine_manager.rs: updated tests to the new contract.

Test plan

• cargo test -p rocm-dash-tui --lib -- --test-threads=1 → 461 passed
• cargo clippy -p rocm-dash-tui --all-targets → clean
• Verified the new 404 guard fails when the /v1 suffix is reverted
• Manual: probe a vLLM on :8000, open dash chat, confirm a completion returns (no 404)
• Manual: serve a managed vLLM recipe, press Esc during readiness wait, confirm modal closes and the server keeps coming up in the background, then chat works

Stacked on #67 (retire-legacy-tui).

[rominf]

Both bugs are real and fixed at the correct layer — nice, focused change. One blocking item plus a couple of things to tighten before this is ready.

Blocking: the 404 fix has no regression test

probed_default_used_when_nothing_configured_but_reachable asserts r.base_url == DEFAULT_CHAT_BASE_URL, which is tautological — both sides are the same constant. I reverted the /v1 suffix locally and the entire llm test suite still passed, so nothing guards the actual fix. The Esc change is well covered, but the headline bug needs a test that fails on revert, e.g.:

assert!(DEFAULT_CHAT_BASE_URL.ends_with("/v1"), "Rig appends chat/completions to base_url");

or assert the resolved base_url ends in /v1.

Please address

• Scope of the Esc change is wider than the description says. The summary lists 6 overlays, but on_console_key is the shared seam for 12. The unlisted ones include onboarding: Esc on a running onboarding step now closes the whole onboarding wizard (onboarding.rs:180, *ob = None). It's consistent with what q already did, but it's an unmentioned behavior change in a guided flow — worth calling out so it's a deliberate decision. (For what it's worth, I checked every caller's Unhandled arm and none branch on Esc, so nothing screen-specific gets silently intercepted.)
• Stacked PR isn't a draft, and CI hasn't run. It's correctly noted as stacked on #67 and targets retire-legacy-tui, but it's marked ready with no checks reported yet, so the test/clippy results are local-only at this point.

Nit

• The doc comment says the new value "Mirrors ... VLLM_ENDPOINT / LEMONADE_ENDPOINT", but those use localhost while this uses 127.0.0.1 — they share the /v1 suffix and port, not the host. 127.0.0.1 is the right pick (unambiguous IPv4 loopback, and it matches the live test in agent.rs); just consider softening the wording to "matches the /v1 convention of ...".

[michaelroy-amd]

Thanks @rominf — all three addressed in a1fa5f7 + a description update.

• Regression test (blocking): probed_default_used_when_nothing_configured_but_reachable now asserts the resolved base_url ends in /v1. I confirmed it fails on revert (reverted the suffix locally → the test panics; restored). No longer tautological.
• Scope: rewrote the description — the shared on_console_key seam has 12 callers, not 6, and I now explicitly call out the onboarding behavior change (Esc on a running step closes the wizard, consistent with q) as a deliberate decision. Thanks for checking the Unhandled arms.
• Nit: softened the doc to "matches the /v1 convention of VLLM_ENDPOINT / LEMONADE_ENDPOINT" and noted those use localhost while this uses 127.0.0.1 (unambiguous IPv4 loopback, matches the live test in agent.rs).

On the CI/draft note: it targets retire-legacy-tui (#67), so checks trail the stack; results are local for now (461 tests + clippy clean). Re-review when you have a moment?