Repo size + boundary gates are blind to first-party Go dirs outside internal/ and cmd/

[sunholo-voight-kampff]

Found at sprint-planning time for #498 Lane B (mission iteration 138). Both findings were
reproduced first-party by the controller after the planner surfaced them, each with a
known-positive control so the zeros are measurements and not broken patterns.

1. make check-file-sizes enumerates find internal cmd only

make/code-health.mk:126 — the gate body is:

for file in $$(find internal cmd -name "*.go"); do

Measured:

Query	Result
find internal cmd -name '*.go' | grep -c 'internal/apiserver/'	45 (known-positive: the gate does work)
find internal cmd -name '*.go' | grep -cE '^(runtime|std)/'	0
find runtime std -name '*.go' | wc -l	6 — the files genuinely exist

So every first-party Go package outside internal/ and cmd/ — today runtime/, std/,
testutil/, scripts/, tools/* — has no 800-line enforcement at all. This is not
hypothetical: #498 Lane B adds a new top-level public package serveapi/, and its design doc's
acceptance criterion "make check-file-sizes passes" would have passed identically if
serveapi/serveapi.go were 5,000 lines. The sprint plan works around it with an explicit
find serveapi -name '*.go' | xargs wc -l assertion, but that is a per-sprint patch of a
repo-wide gap.

2. make check-boundaries iterates three fixed package sets, and apiserver is in none

scripts/check_boundaries.sh enumerates CORE_PKGS, DASHBOARD_PKGS, CORE_SURFACE_PKGS.

Query	Result
grep -c apiserver scripts/check_boundaries.sh	0
grep -c serveapi scripts/check_boundaries.sh	0
grep -c parser scripts/check_boundaries.sh	4 (control — the instrument matches)
bash scripts/check_boundaries.sh	rc=0, OK: no architecture boundary violations.

The gate therefore passes whether or not internal/apiserver (or a future serveapi/) imports the
compiler core directly. The script's own header comment already anticipates this failure class:
it warns that a wrong package value "silently makes the gate pass on every import match".

Why this matters beyond the two gates

Both are vacuous-pass defects: the gate runs, exits 0, and is quoted downstream as evidence for
a claim it never covered. The #498 Lane B design doc had already measured the boundary-gate gap
(its row V18) and still listed make check-boundaries as an acceptance criterion — the measurement
was recorded and the conclusion built on it was not corrected.

Suggested fix

• Derive both gates' file lists from go list ./... (or find . -name '*.go' minus vendor/
  worktrees/node_modules), not from a hardcoded directory pair / package array.
• Add an anti-vacuity floor to each: assert the enumerated file count is above a known minimum and
  exit 1 loudly if the list comes back short, so a future re-scoping cannot silently empty it.
• Add a self-test fixture that a deliberately-oversized file outside internal/+cmd/ is caught.

Not urgent enough to outrank the queue; filing so it is not re-discovered a third time.