`worklog doctor`: environment and connectivity diagnostics

[Saber5656]

Title

worklog doctor: environment and connectivity diagnostics

Summary

Implement worklog doctor: a check matrix over config, each enabled
source, the LLM endpoint, and the output directory, printing PASS/WARN/FAIL
lines with actionable fixes (notably the zsh EXTENDED_HISTORY guidance) and
exiting non-zero only on FAIL.

Context

The reference machine demonstrates the two onboarding cliffs doctor exists
for: zsh history without timestamps, and an LLM endpoint/model that isn't
running/pulled yet. Doctor turns both from silent degradation into guided
setup.

Scope

• internal/cli/doctor.go (replace stub) + tests.

Detailed Requirements

1. Output format (stdout — doctor's report IS its output; stderr only for
   harness errors): one line per check:
   PASS config: loaded (<path>) / WARN zsh: … / FAIL llm: …,
   followed by an indented fix: line for every WARN/FAIL with a concrete
   command or config edit. Final summary line
   doctor: N pass, M warn, K fail. Exit 1 iff K>0, else 0.
2. Checks (in this order; each isolated — one failure never skips later
   checks):
   • config: load + validate (validation errors → FAIL each, fix
     references the key). Unknown-key warnings → WARN each.
   • zsh (when enabled): file exists+readable (missing → WARN, fix:
     check path/HISTFILE); parse first 2000 entries: ExtendedCount==0 →
     WARN with the exact fix
     echo 'setopt EXTENDED_HISTORY INC_APPEND_HISTORY' >> ~/.zshrc
     (new entries only — note that old entries stay undated); else PASS
     with entry count.
   • git: binary found (ErrGitNotFound → FAIL, fix: install git);
     author identity resolvable (else WARN, fix: set [sources.git] authors or git config --global user.email); each root exists
     (missing → WARN); repos discovered (count in PASS line; zero repos →
     WARN).
   • claude-code / codex (when enabled): dir exists (missing → WARN,
     fix mentions the config key), count of session files with mtime within
     7 days (0 → WARN "no recent sessions").
   • llm (when llm.enabled): endpoint parses + static loopback IP check
     (violation → FAIL with ADR-002 one-liner); Models(ctx, 5s timeout)
     reachable (refused → FAIL, fix: ollama serve or start the server;
     ErrNonLoopbackBlocked → FAIL with the blocked host/IP); configured
     model ∈ returned ids (empty model → FAIL listing first 5 available
     ids as the fix; not-in-list → WARN listing near matches). When
     llm.enabled=false: single line PASS llm: disabled.
   • output (when [output].directory set): exists → writable check via
     os.Stat + unix.Access(W_OK)-equivalent (use
     os.CreateTemp? NO — doctor must not write (I3 applies to doctor
     strictly: it's a diagnostic); use permission-bit heuristic +
     Stat-based ownership check and label it WARN (heuristic) when
     uncertain); missing → WARN "will be created on first --save".
3. All checks run with a 5s per-check timeout (hung LLM endpoint must not
   hang doctor).
4. Global flags (DESIGN §4): --config and --verbose are honored;
   --date and --lang parse successfully but are ignored by doctor
   (doctor output is English-only and date-independent; --verbose prints
   a one-line note that they were ignored). --verbose adds detail lines
   (resolved paths, counts) — still stdout.
5. Testability: checks are small funcs returning
   {Status, Name, Detail, Fix string}; the runner formats. Fake
   providers/clients via the same seams as earlier issues (fixture HOME +
   httptest fake for LLM).

Acceptance Criteria

• Fixture matrix test: healthy HOME → all PASS, exit 0; each induced
  failure (missing zsh file, simple-format history, git absent from
  PATH via env manipulation, missing model, endpoint down, non-loopback
  IP endpoint) produces its documented line+fix and correct exit code.
• zsh WARN includes the exact one-line fix command above.
• LLM FAIL on empty model lists available ids from the fake server.
• Doctor performs ZERO writes (tree-hash over fixture HOME, including
  the output dir case).
• Full run with hung LLM (server accepts, never responds) completes in
  < 10s (timeout proven).
• Output is stable/deterministic (sorted, no timing text) — golden
  test for the healthy case (doctor output is English-only per DESIGN
  §4; --lang parses and is ignored).
• worklog doctor --date 2026-01-01 --lang en parses and behaves
  identically to a plain run (ignored-flags contract).

Validation

go test -race ./internal/cli/ -run Doctor -v in PR + healthy-case golden
attached. Manual run on the reference machine encouraged; paste is fine
(doctor output contains no personal content beyond paths — verify before
pasting).

Dependencies

07, 09, 11, 14, 15, 23.

Non-goals

Auto-fixing anything (doctor never writes), model pulling, benchmark/latency
measurement.

Design References

• docs/DESIGN.md §4, §13, §15 (U2/U3), research docs (fix texts)

---

Source of truth: docs/issues/28-cmd-doctor.md (PR #1, branch docs/v1-design). If this issue and the repo docs disagree, the docs win. Execution order and dependencies: docs/ISSUE_PLAN.md (this is issue 28 of 33).