fix(gates): resolve the content screener at generation time; count aliased context files once

[alkisyuv]

Two defects, both cases of a measurement or a gate being wrong about the repo it runs in.

1. Generated hooks guessed at the content screener by path

The emitted content-screen resolution carried a bare [ -x ./dist/cli.js ] check ahead of the
etymd on PATH. That path is simply where a great many CLI projects build, so any repo that
builds its own binary there had the hook invoke that binary as the screener. It does not know
screen, so:

• the commit door failed closed on every commit (error: unknown command 'screen', rc=1);
• the push door, which ignores the screen's exit status by design, skipped the whole-tree
  pass in silence — the worse of the two, since the repo still looks screened.

The trap armed itself on a plain dependency install, because that runs the repo's build. Measured
in three repos that build to that path; a fourth was unaffected only because nothing had built it.

The fix. Resolution is now CONTENT_GATE, then the etymd on PATH — in every repo. The
dev-build arm survives for the repo that develops the screener itself, where gating on an
unreleased build is the whole point, but isSelfBuildRepo decides it at generation time from
the manifest name and emits it nowhere else. It keys on the manifest name rather than
facts.name, which falls back to the directory basename and would re-arm the same coincidence
for any directory that happens to share the name.

Consumer hooks get strictly less shell than before. A repo needing a different runner for one
invocation still has CONTENT_GATE.

Pinned by four tests, including the reported symptom end to end: a repo with a foreign executable
dist/cli.js makes a real git commit that runs the screen through the ambient etymd with no
CONTENT_GATE override, and never reaches its own binary.

2. Context economy counted a symlinked instruction file twice

AGENTS.md symlinked to CLAUDE.md is the ordinary way to serve harnesses that read different
names. Both names were measured separately, which doubled the reported always-loaded footprint —
enough to manufacture an over-budget finding out of nothing — and fired the heavy-file finding
twice for one file.

Candidates resolving to the same inode (symlink or hardlink) are now counted once and reported
under every name they answer to (AGENTS.md → CLAUDE.md), in etymd audit and etymd context
alike. The finding ID stays the primary path so a ledger entry survives a repo adding or dropping
a name, and the merge is stated in the lens disclosures rather than applied silently.

Notes

• Pack v8 → v12. Numbers 9–11 are claimed by unreleased work on the other two open branches; 12
  is above all of them, so any merge order resolves to it.
• This repo's own hooks are regenerated in the same change, and keep the dev-build arm — it is the
  dogfood case.
• Green: 269 passed | 8 skipped, format:check, typecheck, and the self-audit clean at exit 0.
• The self-audit discloses Baseline was approved under pack v8; current pack is v12 —
  re-approving the baseline is a separate call.
• Conflicts with the other two open branches in the four generated hook files and
  src/pack/version.ts; mechanical, as noted there.