feat(gate): scan chosen folders for terms, and project codes outside their folder

[pasrom]

Problem

The gate's leak check scans forbidden terms only in audience: customer docs. A customer name in an
internal doc in a customer-neutral folder passes CI, and that is the slip a team brain most needs
caught. The terms file is gitignored, so CI never has it, and fork PRs get no secrets to supply it.

Change

Three opt-in gate keys. With all three unset, the gate behaves exactly as before.

• term_scan_prefixes (list): folders whose docs are term-scanned whatever their audience.
  bms and bms/ both match bms/..., never bmsx/....
• project_prefix (str): each tracked subfolder under it is a project code. A code in any doc outside
  its own subfolder is a gate-leak:code error. It matches whole words, treating _ as a separator,
  and the code is redacted in the report. Files directly in the project folder, such as its
  _index.md, may list every code. This check needs no terms file, so it also runs in CI, fork PRs
  included.
• project_code_pattern (regex): a subfolder name must fully match it to count as a code. This is for
  a project folder that also holds non-customer topics; without it, a folder named after a regulation
  flags every mention of that regulation. An invalid regex fails fast.

Both path-keyed scans run in the text pre-pass with the secret scan, so they also cover exempt files
and _index.md. Codes come from the tracked files, so a local-only folder never changes the result.
The needles are normalised once. The code regex runs only after a plain substring hit, because its
lookbehind makes a full scan slow (measured about 34 times faster on an 840-doc brain). Per-file paths
are made POSIX where they are built.

Verification

• All four smoke tests pass: gate 50, index 30, team 96, encoding 4.
• Every new behaviour was mutation-checked: scope, prefix slash, hub skip, own folder, tracked-only
  codes, pre-pass coverage, single report, word boundary, _ separator and pattern. Each mutation
  fails at least one test.
• Run against a real team brain, main plus 11 open PRs, with project_prefix and a pattern: no
  false positives. A planted code in a neutral doc was reported and redacted.
• Without a pattern, a topic folder such as projects/cra/ would flag every "CRA" in the brain. That
  case is why project_code_pattern exists.

A release is a separate version bump.