pnpm demo prints its "now create the three requesters" instruction ~290 lines and 120 ERROR lines before the ready banner, so nobody reads it

[zhuangjianguo]

Fixes #49

What changed

scripts/demo.mjs only. The operator instruction — create the three business-requester accounts, run again — used to print inline, immediately before startDemo() handed the terminal to the demo boot. It now prints after that boot has come to rest, and it says what the seed's owner_id errors are while the same reader is still looking at them.

The defect, measured (not repeated from the card)

Clean-database pnpm demo runs (rm -rf .objectstack dist first), captured to a file. Base origin/main @ a7b7db5; branch at 405d67f:

	measured on main	on this branch
operator instruction at line	12	167
✓ Server is ready at line	147	136
terminal comes to rest (Press Ctrl+C to stop)	169	159
boot output after the instruction	157 lines	0 lines
total lines captured	172	195
ERROR lines in the boot	2 (sys_oauth_resource UNIQUE, unrelated to the seed)	same 2
ERROR [SeedLoader] lines	0 in two runs, one held 150s past the banner	0 in a 30s window; 120 in a 90s window

The card's magnitudes are different from mine and the report on #49 says why. The defect is confirmed either way: 157 lines of boot output stood between the instruction and the resting point. The cause the card names is confirmed straight from the seeded database — every clm_contract row landed with owner_id NULL — and from the loader's own error text: Deferred reference UNRESOLVED after pass 2 — clm_contract.owner_id stays NULL … no such sys_user row exists, once per contract row.

The measurement that shaped the fix

The ready banner is where the boot comes to rest. On a first boot it is not always where the seed does:

07:03:46.862  WARN [Seeder] Inline seed exceeded 8000ms budget … continuing in background
07:03:48      ✓ Server is ready · Press Ctrl+C to stop · (this note)
07:05:09.768  ERROR [SeedLoader] … 120 lines, from the background continuation
07:05:09.827  WARN [Seeder] Seed loading completed with … dropped record(s) and … error(s)

82 seconds after the banner. The same command on main, held open for 150s past its banner, never emitted those lines at all while seeding the same corpus — same code, same fixture, two different clocks. Nothing the script can observe tells the cases apart (the WARN, the errors and the summary are all in the child's inherited stream), so the note does not claim the errors are above it. It names the clock that puts them above — seed inside its inline budget, which is what the dogfood pass saw — and the one that puts them below, and it is the frame either way. It is not printed twice to cover both.

Adjudicated by the seat: the strict ordering is a platform gap (nothing an app can observe says the boot has come to rest), filed upstream by the seat and not closed here. No authenticated poll, no pipe.

What the note does NOT print: a census

Earlier revisions of this note printed the counts I had measured — the number of ERROR lines, the per-object row totals, the figure in the loader's summary. Those are out. Every one was true on the boot it was measured on and none is defended by anything, since no gate reads printed prose: #47, in flight in this same batch, deals legal_owner across the seeded corpus and would falsify the error-line count in silence; #41 would move the row census. A stale census printed to an operator as reassurance is the demo telling a stranger a false number about its own health.

So the note states the shape, which is true at any count: one error line per contract whose requester account does not exist yet, every seeded row present with only owner_id NULL, and the loader's summary contradicting the database quoted as its own accounting rather than as a figure. The one number the file still prints is derived at runtime (Math.min(log.length, LOG_TAIL_LINES)), which is the pattern. A three-line rule in the note's doc block says why, so the census does not come back.

How the note reaches the resting point

stdio: 'inherit' is kept and the boot's own output is untouched — not filtered, not reformatted, not dimmed. Piping the child would give an exact "it stopped printing" signal at the cost of the child's TTY, and a boot behind a pipe loses its colour: the seed's ERROR lines would have arrived dimmed by the very edit meant to explain them. So readiness is a TCP probe on the port os dev will bind — resolved the way os dev resolves it (--port, then OS_PORT, then PORT, then 3000) — plus a 2.5s settle for the 22-line banner tail. Both blind cases degrade to the pre-fix behaviour (a note printed mid-stream), never to a lost note; a boot that dies gets no note at all, because os dev has already said why.

Before / after, as a reader experiences it

Before (origin/main, last lines on screen — the instruction is 157 lines above this):

  ⚠ Boot diagnostics — 1 warning logged during startup:
    WARN [Seeder] Inline seed exceeded 8000ms budget for app.objectstack.hotclm; continuing in background…
    run with --log-level debug to watch the boot stream live

  Press Ctrl+C to stop

After (this branch at 405d67f, last lines on screen, copied from the capture):

  ────────────────────────────────────────────────────────────────────────
   Before you open the app — one setup step, and one thing about the log
  ────────────────────────────────────────────────────────────────────────

  1. The seeded contracts have no owner yet. Every contract is launched by
     one of the three business requesters DESIGN.md §10 asks you to create,
     and no seed may create a user (§10) — so until those accounts exist,
     `owner_id` is NULL on every contract row and 我的合同 stays empty. Add
     them in Setup → Users and run this again; the README names them and
     says who gets what.

  2. The `ERROR [SeedLoader]` lines are that same missing owner: one per
     contract whose requester account does not exist yet, expected on a
     first boot, and nothing is lost to them. The loader defers `owner_id`,
     finds no such account on its last pass, writes the row anyway and
     leaves that one column NULL — every seeded row is in the database and
     only its owner is missing. Step 1 is what fills it in.

     The loader then signs off with a summary that counts those rows as
     dropped while they sit in the database: its own accounting, not the
     state of your data, and upstream as objectstack#17177. And where those
     lines fall relative to this note is a clock, not a verdict — above it
     when the seed finished inside its inline budget, up to a couple of
     minutes below it when the boot said `WARN [Seeder] … continuing in
     background`.

Gates — exit codes captured before any pipe

Re-run on the reworked tree (405d67f), each gate redirected to a file first and $? read after, so no exit code passes through a pipe:

GATE validate       EXIT=0
GATE lint           EXIT=0
GATE typecheck      EXIT=0
GATE lint:i18n-gate EXIT=0

Plus node --check scripts/demo.mjs → 0, and a control-character self-scan of the file (grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]') → no matches.

Acceptance notes

• ⛔ No error line suppressed, downgraded, filtered or reformatted. The note explains them; it does not hide them. The runs above show the same loader errors and the same 2 boot errors as main.
• ⛔ README.md (README.md opens with Status: M0 — scaffold and configuration domain on an M1/M2/M3-complete tree, and its setup section never says the position assignment binds by name #46), DESIGN.md (PR Record four maintainer rulings in DESIGN.md §03, §06 and §09 #40) and src/ untouched. Diff is scripts/demo.mjs only.
• No changeset (this repo has no changeset gate).
• No browser evidence, and none is owed: this changes terminal output only. No object, view, page, app, flow or action moved, and there is no runnable surface to drive. The reader-facing surface is pnpm demo's stdout, and it is measured above.
• Noted, not filed — the note keeps the pre-existing sentence's 我的合同, a zh-CN view label inside English default-locale output. Preserved verbatim rather than rewritten under a card whose surface is print order. Follow-up owner: [Decision] pnpm demo 开箱即用时没有任何账号能打开「我的合同」——要不要让脚本建账号，与 #11 第一问耦合 #28 (same text), or README.md opens with Status: M0 — scaffold and configuration domain on an M1/M2/M3-complete tree, and its setup section never says the position assignment binds by name #46 if the README half is localised first.
• Noted, not filed — the boot logs 2 unrelated ERROR lines on every clean start ([sql-driver] DATABASE_ERROR plus an sys_oauth_resource UNIQUE constraint insert failure). Identical on main, so not caused here, but a reader counting red lines counts these too. No queued card touches that path.

🤖 Generated with Claude Code

https://claude.ai/code/session_01R3n3GGzobdegM4HUzah1iR