test(guard): tests must not gate on paths outside the workspace

[noahgift]

The monorepo consolidated 20 sibling repos in-tree. Tests written before the
merge still gate on the OLD sibling checkouts:

let has_q4k = file_exists("/home/noah/src/realizar/src/quantize.rs");
if !has_q4k { eprintln!("SKIP - realizar not found"); return; }

realizar IS crates/aprender-serve now. That path cannot exist again on any
machine, so every test behind such a gate is permanently and silently green.
Five of them remain, in falsification_2x_ollama_tests.rs and
falsification_correctness_tests.rs, and one runs a command with
.current_dir("/home/noah/src/realizar"). Not one reads the file it probes --
the path exists only to decide whether to skip. F102's gate is purely
decorative: its body times two inline loops and never touches realizar at all.

They are invisible twice over: workspace-test runs --lib, so these integration
targets are never compiled either. A skip nobody sees, in a test nobody runs,
behind a claim somebody trusts.

Scope: 85 out-of-workspace paths across 1921 test files. Includes gates on this
repo's OWN pre-monorepo layout (/home/noah/src/aprender/src/format/mod.rs --
that file is crates/aprender-core/src/format/ now).

This is a RATCHET, not a cleanup. Repointing the existing gates makes dormant
tests execute for the first time and will surface real failures; that is its own
change with room for the fallout. This stops the population growing and makes
the debt a number that can only fall.

Verified:

• case table, 3 defect shapes + 4 legitimate paths (workspace-relative, /tmp,
  env var, relative). The table caught my first version being BLIND -- the
  find pruned tests directories, excluding the only thing it must scan.
  Fifth time a guard regex in this repo was wrong; fifth time a table caught it
  and review would not have.
• mutation: add one new out-of-workspace gate -> RED, listing every violation.
  Restored -> PASS.
• vacuity: fails if fewer than 200 test files are scanned, so a broken scan
  cannot report clean.

Refs #2474

Co-Authored-By: Claude Opus 5 noreply@anthropic.com