docs: re-derive the os compile / os validate transcripts from real runs

[os-steve]

Fixes#9092

#4409 collapsed the per-concern authoring checks into one registry, so the three
authoring commands print one step for the whole table. Two sample transcripts still
reproduced the pre-#4409 step list. Both blocks are re-derived from real runs, not
hand-edited — retired steps are replaced by what a run actually emits, never deleted,
because a transcript is a reproduction of program output and deleting lines manufactures
output just as surely as leaving stale ones does.

How the transcripts were produced

CLI built from this branch, then run against two real projects:

block	project	command
cli.mdx · os compile	the blank starter scaffolded from packages/create-objectstack/src/templates/blank	node packages/cli/bin/run.js compile
build-with-claude-code.mdx · os validate	a Support Desk rebuilt from this page's own step-3 listings (object + action + views + a Support-nav app)	node packages/cli/bin/run.js validate

The failing run is reproducible rather than remembered: the fixture carries the page's own
bare-reference predicate, visible: 'status != "resolved"', on resolve_ticket. Flipping
that one string to record.status != "resolved" && record.status != "closed" turns the
same fixture into the clean run — so both blocks on that page come from one project in two
states, which is exactly what the page narrates.

Cross-checked against the sources afterwards, never derived from them: validate.ts has
five printStep calls and compile.ts has twelve, several conditional. Grepping
printStep would have produced a transcript no run emits — the same defect this card
exists to fix, reached from the opposite direction.

What changed

cli.mdx · os compile — five retired step lines replaced by the two a run prints:

 → Validating protocol compliance...
- → Validating expressions (ADR-0032)...- → Checking dashboard widget bindings (ADR-0021)...- → Checking dashboard action references (ADR-0049)...- → Checking SDUI styling (ADR-0065)...- → Checking security posture (ADR-0090 D7)...+ → Running author-time rules (41)...+ → Checking capability providers (#3366)...
→ Collecting package docs (ADR-0046)...

build-with-claude-code.mdx · failing os validate — four defects, one more than the
card listed:

1. → Validating expressions (ADR-0032)... ⇒ → Running author-time rules (41)...
2. ✗ Expression validation failed (1 issue) ⇒ ✗ Author-time rules failed (1 issue)
3. the rule: expression-invalid at ... trailer was missing entirely — added
4. ✗ EEXIT: 1 is not emitted at all and was removed. Not in the card. Verified both
   piped and under a real TTY; isExitSignal in packages/cli/src/utils/format.ts:131
   is what swallows it, and migrate/meta.ts:533 carries a comment about suppressing
   exactly this line. The ✗ Author-time rules failed header already carries the failure
   signal, and the run still exits 1.

build-with-claude-code.mdx · clean os validate — the blank line validate.ts:297
prints unconditionally between the manifest line and the stats.

Two measurements that correct the card

The source: line is correct and was kept. The card predicted findings now carry "a
hint line and a trailer ... rather than a source: line". Measured: source:is
this rule's hint, so it survives verbatim; what was missing is the rule: ... at ...
trailer underneath it. Had the card's reading been applied literally, a correct line
would have been deleted.

The missing blank line exists on one page, not two. The card and the dispatch both
scope it to "both pages". On cli.mdx both candidate blocks already have it —
os compile prints ✓ Build complete then a blank line then Data:, and os info
already has the blank line after Type: app. Only build-with-claude-code.mdx was
missing it. The instruction stands; one of its two targets did not exist.

Sweep of the other os * output examples on cli.mdx

Every transcript-shaped block on the page was enumerated and checked, not just the one the
card named. os info was re-run and its structure matches a real run line for line —
no drift, no change needed. The os migrate plan, database-in-use, and os login --json
blocks carry no step list and are unaffected by the registry collapse. So os compile was
the only block on the page with real drift.

What was deliberately NOT changed, and why

The stats lines and timings. Every line I changed is a project-independent fact about
the CLI, provable from any run. The stats lines are fixture identity, and this page's
Support Desk cannot be reconstructed exactly — the page labels its own code listing "a
representative result" and shows 4 fields, while the transcript says 6. Pasting my
reconstruction's 1 Objects 4 Fields would have substituted my fixture's identity for
the page's and, worse, broken agreement with
content/docs/deployment/validating-metadata.mdx, which says 1 Objects 6 Fields, is
out of scope here, and was corrected by #9034 earlier today. Reverse-engineering two extra
fields to hit "6" would have been manufacturing. So the numbers are left, and the
discrepancy is filed rather than guessed at — see below.

Block boundaries. The failing block remains an excerpt starting at the first step line
(the Loading configuration / Config: / Load time: preamble was already elided before
this PR). Expanding it would have meant importing a Load time: value from elsewhere or
publishing my machine's, so the pre-existing boundary was kept.

The cross-reference this repairs

validating-metadata.mdx sends readers here "for the bare-reference example verbatim".
That word was load-bearing and false. The finding block now matches a real run, so it is
true again. The corrected block does not break the cross-reference in the other direction:
validating-metadata.mdx shows the same header vocabulary, the same registry step, the
same 41 count, and already has the blank line.

Volatility note (raised, not acted on)

Running author-time rules (41) encodes a live count that will go stale again — the same
class of defect as #9034. There is a way to make the block not encode it, but every option
changes the card's shape, so per the dispatch this is flagged rather than acted on
unilaterally.

Verification

Gate union re-run after the final commit, at 689551a31:

check:cross-package-test-inputs PASS
check:docs-audit-scope PASS
check:docs-redirects PASS
check:role-word PASS
check:nul-bytes PASS

The list was re-derived from the actual changed paths via
node scripts/pm/dispatch-gates.mjs, which named the same five as the dispatch — no
additions. Docs-only; no published package changes, hence skip-changeset.

---

Generated by Claude Code