fix(hermes): surface dashboard launch failures with a real diagnostic block

[la14-1]

Summary

• startHermesDashboard() previously hid every failure behind a generic "Hermes web dashboard failed to start — TUI still available" warning. When users reported breakage (e.g. issue [Bug]: hermes on digitalocean fails — hermes dashboard reported as not a command #3407 — Andrew's "hermes dashboard isn't a command" / "the dashboard command of hermes also changed to something else") we had no signal to debug from.
• Now the deploy script installs an EXIT trap that fires on any non-zero exit and dumps a banner-delimited diagnostic block to stderr (which is inherited live to the user's terminal):

  ──── Hermes dashboard diagnostic ────
  hermes binary: /home/.../venv/bin/hermes
  hermes version: Hermes Agent v0.13.0 (2026.5.7)
  hermes subcommands: includes "dashboard"
  ─── /tmp/hermes-dashboard.log (last 30 lines) ───
  <actual hermes process output>
  ─────────────────────────────────────
  

• The trap is cleared on the success path so it stays quiet when things work.
• The TS-side warning now also includes the underlying error message and points users at the diagnostic block for bug reports.
• Bumps @openrouter/spawn to 1.0.45.

What this catches

Failure mode	Old output	New output
hermes not in PATH	generic warning	hermes binary: <not found in PATH> + tail-of-log
dashboard subcommand genuinely renamed/removed	generic warning	hermes subcommands: "dashboard" NOT in --help output
fastapi/uvicorn lazy-install fails	generic warning	full hermes stderr including the import-error message
Web dist build fails	generic warning	the build-error stack tailed from /tmp/hermes-dashboard.log
Port 9119 already taken / bind failure	generic warning	hermes's own bind error from the log

Test plan

• bunx @biomejs/biome check src/ — clean (203 files)
• bun test — 2207 pass / 0 fail (was 2204; +3 new assertions)
• New tests pin: trap '_dashboard_diag' EXIT, the diagnostic banner, all probe commands, trap - EXIT on success, and the failure-surfacing TS-side message.
• Generated script syntax-checked with bash -n
• Subcommand probe (hermes --help | grep -q '^[[:space:]]*dashboard') verified against a live install of hermes-agent v0.13.0 (2026.5.7) from main — matches, so genuine missing-subcommand reports will be loud, not noise.

Refs #3407 — won't close it; that issue stays open until we get a real diagnostic from the next user who hits this and can report the actual cause.

[la14-1]

Note: This PR and #3410 both address #3407 but are complementary, not duplicates:

• fix(hermes): gate dashboard launch behind capability probe #3410: Capability probe — skips dashboard launch if hermes dashboard subcommand is missing.
• fix(hermes): surface dashboard launch failures with a real diagnostic block #3413 (this PR): Diagnostic surfacing — EXIT trap dumps debug info (binary path, version, subcommand list, log tail) when dashboard fails for other reasons.

Recommend merging #3410 first, then this PR. #3413 builds on #3410's version bump (1.0.44 → 1.0.45), so landing #3410 first avoids a version conflict in package.json.

-- refactor/pr-maintainer

[la14-1]

Reviewed both PRs for #3407. Closed #3410 (probe-only approach) in favor of this one — the diagnostic block is the right call because:

1. Silent skip (PR fix(hermes): gate dashboard launch behind capability probe #3410) hides the root cause; this PR surfaces it
2. The trap-based diagnostic dump gives users and us everything needed for debugging: binary path, version, subcommand availability, and log tail
3. Includes 3 new test assertions pinning the diagnostic behavior
4. Version bump included

All CI checks pass. Ready for maintainer review and merge.

-- refactor/code-health