fix(skills): keep the seam gate's preview attached so --project works again

[jrusso1020]

What

Pass --foreground in the seam gate's default preview command, so seam-gate.mjs verify --project starts an attached preview again.

Why

seam-gate verify --project has been failing with preview server exited early since #3310. The film is fine; the gate is misreading a healthy launch.

#3310 made hyperframes preview choose its launch mode from the TTY:

// packages/cli/src/commands/preview.ts:613
if (!options.foreground && !options.interactive) return "background";

seam-gate.mjs spawns the launcher with stdio: ["ignore", "pipe", "pipe"], so interactive is always false and it silently takes the background path. Two things then break:

1. The launcher exits 0 as soon as it has handed off to the managed background server, which is before the server is serving. The readiness loop's child.exitCode !== null check treats that as a dead server and throws.
2. spawnDetachedPreview uses detached: true + unref(), so the server is its own process-group leader. The gate's cleanup kills the launcher's group (process.kill(-child.pid, ...)), which no longer reaches it — every --project run leaks a preview server.

The misdirection is the expensive part: the message says the film's preview died when the launch actually succeeded.

How

One flag on the default --server-cmd. --foreground is the CLI's own supported opt-out, added by #3310 for precisely this caller — its help string is "Force an attached preview in a non-interactive shell."

That restores both invariants the surrounding code already assumes: a non-null exit code means the server really died, and the server stays inside the process group cleanup kills. So it fixes the false failure and the leak with the same change.

Deliberately not doing what looked like the obvious alternative — making preview foreground-by-default again. Backgrounding in every launch mode is what #3310 set out to do, and the opt-out for non-interactive callers already exists. The gate was the thing that hadn't been updated.

Both mirrored copies (.claude/skills/ and .agents/skills/) are updated; scripts/check-skill-mirror.mjs passes byte-for-byte.

Test plan

The causal chain is pinned by evidence rather than inference:

• The gate's spawn really is non-interactive. Spawned a child with the gate's exact stdio: ["ignore","pipe","pipe"] shape: stdin.isTTY=undefined stdout.isTTY=undefined => interactive=false.

• Non-interactive without --foreground resolves to background. Already pinned by this repo's own committed table in packages/cli/src/commands/preview.test.ts:146 — {background:false, foreground:false, interactive:false, ...} → "background".

• --foreground escapes it even when non-interactive. Same table, preview.test.ts:166 and :176 — {foreground:true, interactive:false, ...} → "dev" / "local", both attached modes.

• node --check on both copies, oxfmt --check clean, check-skill-mirror.mjs passes.

• Unit tests added/updated — no new test; the behaviour this depends on is already pinned by the existing previewLaunchMode table, and the changed line is a default command string with no test harness around it.

• Manual testing performed

• Documentation updated (if applicable) — the --project contract is unchanged, so references/seam-gate.md still describes it correctly.

— Rames

[miguel-heygen]

Exact-head review at 947f1a38543c58a2db092ca0ca649c55761c8d08.

The fix is at the correct boundary. motion-doctrine/scripts/seam-gate.mjs:74-78 opts this known non-interactive caller into the CLI’s existing attached mode instead of undoing #3310 globally. That preserves the readiness invariant (child.exitCode means the serving process group actually died) and keeps the foreground preview under the detached launcher group that cleanup already signals. Both skill mirrors are byte-identical.

I traced the dependency rather than taking the body: previewLaunchMode maps non-interactive/no-flag to background and explicit foreground to attached dev/local modes; the managed background child is detached/unref’d, so the old gate cleanup could not reach it. The one flag closes both the false early-exit and leaked-server mechanisms.

No blocking findings. Current-head CI is fully green; the relevant Test: skills, skill mirror/lint, CodeQL, format, and regression contexts all pass.

Verdict: APPROVE
Reasoning: The change uses the supported caller-level opt-out, restores the gate’s existing process/liveness assumptions, and leaves global preview behavior untouched.

— Magi