fix: route interactive diagnostics away from TUI output

[sam-saffron-jarvis]

Summary

Follow-up to merged #1183: keep application diagnostics off renderer-owned stdout/stderr in chat and the interactive skills, MCP, sessions and theme views. The scoped runtimeoutput sink writes application diagnostics to restricted $XDG_DATA_HOME/term-llm/diagnostics/tui.log (or the usual home data fallback); actionable session warnings still use the chat footer. It does not replace global descriptors or default loggers, and ordinary CLI, service, renderer, explicit terminal hand-off and auth-login output remain unchanged.

Additional follow-up at 894ac0bc addresses the completed claude-bin:opus-high review (verdict REQUEST CHANGES):

• B1 resolved: Claude/Grok CLI failure and prompt-truncation retry logs now use runtimeoutput.Error/Info with their original severity and fields. A scoped Go AST guard detects future direct slog, log.Print*, stdout/stderr writes in interactive worker packages and selected chat-reachable command files; named exceptions are documented for guarded authentication, CLI-only output, and unrecoverable forced exit. It intentionally does not flag source string fixtures or renderer/auth output wholesale.
• B2 resolved: Grok missing and invalid/expired credentials refuse device-code prompts while an interactive UI is active, reporting term-llm auth login grok; missing and expired credential paths have regression coverage.
• B3 resolved: ChatGPT refresh failure status no longer prints to stderr while an interactive UI is active; the existing guarded prompt returns the actionable login error instead.
• M1 resolved: failure to open diagnostics is no longer a prerequisite for starting an interactive UI. One startup warning is emitted before renderer ownership, then diagnostics go to an explicit discard sink for that UI. A genuinely overlapping UI still errors. No CWD diagnostic fallback is used when HOME/XDG data cannot provide an absolute path. Tests cover warning delivery, suppression, overlap and sink cleanup.
• M2 resolved with documented limit: runtimeoutput.Printf writes raw formatted bytes rather than slog-escaped messages. Debug sections and analogous provider CLI/request blocks now write atomically, preserving readable multiline output. User-facing debugging docs distinguish this file from session debug-log JSONL and warn that raw requests/tool results can contain sensitive data; existing debug flags still gate that output. No automatic rotation/retention is introduced; inspect and rotate/delete the file as needed. No new size/time thresholds or flags were added.
• M3 resolved: child-run MCP/approval progress uses the diagnostic writer rather than footer notices; actionable session-store warnings retain the footer, and child failures remain in run results. A regression test checks progress does not toast. Termhost lifecycle debug also routes safely during UI ownership. Chat diagnostic ownership now follows actual terminal renderer output, including chat --auto-send with redirected stdin; reload/terminal-host gating remains separate.
• The tracked audit records intentional direct output and verification limits. skills validate --all includes malformed discovered manifests while keeping first-valid precedence (a shadowed invalid duplicate does not replace a valid skill). During explicit handover/reload there is no active UI, so ordinary stderr is expected. A same-user attacker who can replace the diagnostic file between Lstat and OpenFile remains an accepted TOCTOU boundary for now; the directory is private (0700) in the normal case. No platform-specific no-follow implementation or descriptor-wide redirect was introduced.

Verification

At follow-up head 894ac0bc:

• make build, go build ./..., go vet ./..., make complexity, git diff --check: passed.
• Isolated go test ./... using temporary HOME/XDG and preserved GOMODCACHE: passed after follow-up changes.
• Focused go test -race ./internal/runtimeoutput ./internal/llm ./internal/termhost ./cmd -run 'Test(InteractiveWorkersDoNotBypassDiagnosticSink|InteractiveDiagnosticsOpenFailureWarnsBeforeUIAndDiscards|ChildProgressStaysOutOfFooter|GrokPreflightCannotPromptInsideTUI|AuthPromptsRefuseActiveTUI|DebugSectionsRemainRawAndAtomicInInteractiveDiagnostics|BestEffortDiscardOnOpenFailure|RawDebugBlocksAndStructuredLevels|OldDiscardCloserCannotReleaseNewUI|LifecycleDiagnosticUsesActiveSinkInsteadOfTerminal|LifecycleDebugReportsSafeRateLimitedErrorClasses|ChatSessionWarningsAcrossSwitchDetachAndClose|ChatOutputOwnershipIndependentOfInput)' -count=5: passed. A lifecycle test's close-vs-second-state race was observed in one full suite run; the test now accepts the extra safe cancellation class on shutdown, and isolated full suite plus focused repeated/race tests pass.
• Follow-up CI run 36334162751: all six jobs passed (test, race, frontend, cross-build, passkey smoke, changes).
• One non-isolated focused cmd run failed tests depending on the host's extra skills; isolated HOME/XDG passed. Prior PR verification included a real PTY config theme run and CLI validation of valid/invalid skill manifests; no new PTY test was run for this follow-up.

Review status

Ready for review. The completed claude-bin:opus-high review requested changes; the follow-up addresses B1–B3 and M1–M3 with the explicit accepted limits above. Parent integration checks verified the routing/auth/startup changes, clean worktree, matching PR head, passing six CI checks and mergeability. There was no second model review or fresh Opus sign-off on the fix commit; do not interpret addressed findings as an Opus approval. No deployment performed.