✅ Cover projected content's lifetime with real modifiers

[taras]

Why

#203 asks that content projected into a component belong to that component's
resource scope, and lists what has to be covered. The engine already does it —
withInvocation (§4.4), the content scope, the three-stage teardown — and the
specification and website already describe it. Three of the issue's coverage
points had no test:

• a projected persist eval resource stopping with the invocation,
• a projected daemon stopping with the invocation, and
• an executable Markdown test where the ownership is observable without
  JavaScript.

persist and daemon were exercised only in a component's own body (L3, Q5),
and the ordering claims rode entirely on the stub watch modifier in Tier O.
Spec §6.11 promises that a daemon started inside <TempDir> stops before the
directory is removed; nothing checked it.

What changes

No behavior changes. Three tests and their two conformance rows.

Before: the caller's side of the boundary was covered only through the Tier O
stub. A regression that anchored projected persist/daemon resources on the
invocation's eval scope instead of the content scope passed every suite.

After: it fails L8 and the TempDir document.

How it works

L8 (packages/core/tests/eval-persist.test.ts) puts a real persist eval
block on both sides of the boundary. The caller projects one; the component
retains its own after <Content />, which is the order a single scope would
release in reverse. A later sibling reads the timeline:

start:projected, start:own, stop:projected, stop:own

Q14 (packages/core/tests/daemon-integration.test.ts) is Q5 with the daemon
written at the call site and the component body being nothing but <Content />.
It fails closed at both ends: RUNNING is printed only once the process has
been seen alive from inside the component, and the probe after the component
chooses STOPPED only against a pid file that exists — a missing one reports
NOPID. A daemon that never started therefore cannot satisfy the premise.

TempDir.test.md tells the same story in Markdown with no JavaScript. The
daemon traps SIGTERM and records whether its working directory still exists at
the moment it is signalled, writing into a second <TempDir /> that outlives the
first:

<TempDir as="report" /> → <TempDir> daemon traps TERM → invocation ends
  → ALIVE (signalled while the directory existed) + REMOVED (directory gone now)

Review guide

Start with: packages/core/tests/eval-persist.test.ts — L8

Then review:

1. packages/core/src/components/TempDir.test.md — the reader-facing form of
   the same contract
2. packages/core/tests/daemon-integration.test.ts — Q14
3. specs/executable-mdx-spec.md — the L8 and Q14 conformance rows

Look carefully at: the daemon in TempDir.test.md arms its TERM trap
before writing the marker the following block waits on, so the wait cannot
succeed against an unarmed trap.

What must stay true

• Projected resources stop before the component's own — enforced by stage 1 of
  withInvocation's teardown and by the content scope's parentage, checked by
  L8 and by ALIVE in TempDir.test.md.
• Nothing a projection starts outlives the invocation — checked by Q14.

How to verify it

Each test was written against a mutation of
Invocation.useContentScope() that returns the invocation's eval scope instead
of creating the content scope — the pre-#203 anchoring:

• L8 proves boundary ordering and fails if projected resources share the
  component's scope. Under the mutation it produces
  start:projected, start:own, stop:own, stop:projected — the exact inversion.
• TempDir.test.md proves the same thing without JavaScript and fails under
  the same mutation (REMOVED is absent).
• Q14 proves a projected daemon does not leak past the invocation, and fails
  if projected processes anchor at the document scope. Removing its daemon block
  makes it report NOPID and fail on RUNNING, so it cannot pass for a daemon
  that never started. It does not discriminate the content-scope mutation,
  and no test of its shape can: under that mutation the daemon anchors on the
  invocation's eval scope, which is still torn down at stage 3 of the same
  invocation, before the probe runs. Only the relative order changes, and a
  lone daemon has nothing to be ordered against. L8 and TempDir.test.md carry
  that direction.

deno task test
./dist/xmd test packages/core/src --raw

Scope

Included

• Coverage for the three untested points of Bind projected content resources to the component invocation scope #203.
• Conformance rows L8 and Q14.

Intentionally unchanged

• The engine. withInvocation, projection.ts and expand.ts are untouched.
• Spec and website prose. §4.4, §6.3.4 and §6.11 already describe this ownership
  model; §6.11 already states the <TempDir> claim the Markdown test now proves.
• Q5's assertion that a signalled subprocess is reaped stays out of scope, as it
  was in Q7.

Generated or mechanical changes

None.

Scope confirmation

• Every changed file supports the purpose described above.
• Unrelated cleanup and formatting changes are excluded.
• Generated or mechanical changes are clearly identified.
• The description matches the final diff and test results.

[github-actions]

PR #326: ✅ Cover projected content's lifetime with real modifiers

4 files, +135 / -0

Scope

✅ PR scope looks good.

Structural

✅ No structural bloat detected.

Slop

✅ Slop indicators look low.

Static Analysis

✅ Oxlint found no issues.

Correctness

No extraneous code patterns detected.