No per-job read surface for sealed manual jobs: add ob job history

cmd/ob/job.go と cmd/ob/schedule.go から始めて、コマンドと finite_envelope の出力パターンを追跡し、次に internal/journal/journal.go、internal/journal/job_result.go、internal/engine/job.go を読んで、記録されたジョブ実行データを確認します。JobRunRequest とジョブ開始レコードを、PruneCandidates による保持も含めて追跡します。ob job history がジョブごとに指定された結果とメタデータを提供し、不完全な実行を処理し、どのスケジュール実行が ob schedule history に残るかを文書化できれば完了です。

The problem

The two job-execution paths have inverted read surfaces: the lower-risk one is
well instrumented and the higher-risk one is not.

A scheduled job has a full per-job read surface:

• ob schedule history <job> — the newest run records, newest first
• ob schedule logs <job> [run] — the output of one exact activation
• ob schedule list — timer state, next elapse, last trigger

A sealed manual job — the path that exists precisely because the job
declares data_effect: migration or destructive — has neither. ob job
carries exactly two subcommands, plan and run (cmd/ob/job.go:57). After
running one, the only way to see what happened is ob audit, which is
journal-wide and unfiltered: one row per invocation across every operation
kind, with -n as the sole control (cmd/ob/commands.go:219). There is no way
to ask "what has catalog-refresh done, and when."

So the job whose failure matters most is the one an operator can least easily
review after the fact.

It is worse than a clean split between the two paths, because a single job can
land on both. Only deploy_lock: pinned forces data_effect: none
(internal/app/validate.go:450); a scheduled job on the default exclusive
lock may declare migration or destructive, and its timer fires it on cron
like any other. What scheduleRun refuses is the operator-initiated run of
such a job (internal/engine/schedule_run.go:57), redirecting it to
ob job plan and ob job run. So that job's unattended firings are visible in
ob schedule history, its hand-triggered runs are not visible anywhere
job-scoped, and nothing shows both in one place — for exactly the jobs where
the interleaving matters most.

Proposed change

Add ob job history <job> — and, if it falls out cheaply, ob job logs <job>
— mirroring the ob schedule equivalents.

This is mostly a read over data that already exists, but not entirely — see
"What is missing" below.
Every job run is already journaled through the canonical writer, and
JobResultEvidence (internal/journal/job_result.go:8) already records the
normalized, redaction-safe result per invocation:

type JobResultEvidence struct {
    SchemaVersion   string
    Changed         bool
    Provider        string
    BeforeRevisions []string
    AfterRevisions  []string
    Digest          string
}

It hangs off the journal record as job_result
(internal/journal/journal.go:68), and the start record already carries the
operator, timestamp, git SHA, approval class, approved-by, and the migration
backup mode with its override operator and reason
(internal/journal/journal.go:152-178).

What is missing

Three fields a review row wants are not journaled today, so this needs a
small write change before the read command:

• Plan digest. Neither JobPlan.PlanDigest
  (internal/onebox/job_plan.go:45) nor Operation.PlanDigest
  (internal/onebox/operation_types.go:194) ever reaches a Record. The
  journal carries only ApprovalDigest — the grant's digest over a struct that
  contains the plan digest, so it is not recoverable from it. This matters
  most for the interactive ob job run <id> path, which writes neither a plan
  nor a grant to disk (cmd/ob/job.go:223-244): the journal is that run's only
  durable record.
• Release. Present only as free text, Detail: "release=" + current
  (internal/engine/job.go:85). Parseable, but fragile.
• Data effect at run time. Not journaled. A break-glass review wants it.

Suggested: add ReleaseID, PlanDigest and DataEffect to journal.Record,
set them in the job start record, and thread PlanDigest through
JobRunRequest. Keep the existing Detail: "release=" for older readers.

Note also that journal.Summarize will not do the reduction — Started,
Operator, StartedAt and Finished are all gated on Phase == "deploy"
(internal/journal/journal.go:485-506) — so this needs a small job-specific
reducer rather than reuse.

Retention

PruneCandidates (internal/journal/journal.go:315-351) puts job_run
journals in a single "auxiliary" window shared with exec, schedule-run,
schedule-pause and service-apply, sized RetainReleases*2 (default 5, so
10 — internal/engine/deploy.go:572, internal/app/defaults.go:25-26), pruned
on every deploy. So a burst of ob exec can evict the break-glass job run
somebody wanted to review, and ob job history would silently have nothing to
show. Either the command's help text states this plainly, or PruneCandidates
grows per-OperationKind windows. Worth deciding with the command rather than
after it. (ob schedule history is unaffected — its depth is journald's.)

New command, not an ob audit --job filter

Settling the alternative this issue originally left open: a new sibling
command, for four reasons.

1. auditRows deliberately flattens to eleven generic fields
   (internal/engine/audit.go:104-116). Service, JobResult,
   ApprovalClass, MigrationBackup and the release do not survive it. A
   --job filter means adding those to AuditRecord — empty on every non-job
   row — plus a job-specific branch. That is a job reader living inside
   audit's name.
2. ob audit is documented and cross-referenced as the who-did-what
   invocation table, and is the next command for operation_failed and
   cancelled. Widening its shape has blast radius; a sibling has none.
3. ob audit reads N+1 round trips, one cat per journal
   (internal/engine/audit.go:66-81). A job reader can use journal.Journals
   in one round trip and skip every id without the -job_run- infix that
   newOperationID embeds (internal/onebox/service.go:60), never parsing a
   deploy journal at all.
4. Symmetry with ob schedule history (cmd/ob/schedule.go:86-126), same
   finite_envelope output class.

Row shape

Suggested columns, mirroring ob schedule history where the fields allow it:
STARTED OUTCOME DURATION OPERATOR RELEASE EFFECT APPROVAL BACKUP CHANGED OPERATION. Outcome is succeeded | failed | incomplete — a start with no
finish is a crashed run and should say so rather than vanish. The help text
should also say that timer firings of a job which also declares a schedule
live in ob schedule history, not here.

One naming note: ob schedule history's help calls an ob schedule run a
"manual run" (cmd/ob/schedule.go:89), while the code and docs use "manual
job" for when: manual. ob job history should avoid the phrase.

Current workaround

ob audit -n <large> and read past every unrelated record, or
ob audit --output json and filter the journal client-side. Neither is a
per-job view, and both require the operator to know the journal's shape.

Worse: ob audit does not currently render sealed job runs correctly — it
reports them as action job, outcome deployed, with no job name — so the
documented fallback does not actually tell you which job ran. Filed separately;
that bug stands regardless of this issue.

Scope and safety

No change to the one-application, one-host scope. Read-only: no lock, no
fence, nothing written, in line with ob status, ob audit and the
ob schedule read commands. Redaction is already handled upstream —
JobResultEvidence is normalized and redaction-safe by construction, so this
surfaces existing evidence rather than widening what is recorded.

Related: #164 covers the interactive approval prompt on the same ob job path.