[Architecture]: two modules both claim to own "private-looking text"

[JunZ-Leo]

Problem / 问题

Two modules each document themselves as the canonical answer to "is this text private?", and
neither dominates the other.

1. loopx/public_safe_text.py opens with "Canonical private-looking-text rules for public-safe
   control-plane fields", explains that its rule set "used to be copied into each owner, and the
   copies drifted", and names tests/fixtures/public_safe_text_corpus.json as the corpus pinning
   the Python and TypeScript owners to one contract. Consumers: feedback, authority,
   boundary_authority, control_plane/goals/artifact_lifecycle.
2. loopx/control_plane/runtime/public_safety.py carries an independent
   SECRET_LIKE_SURFACE_PATTERN + LOCAL_PATH_SURFACE_PATTERN, used by
   validate_public_safe_value — the guard on the busiest public-output path (12 production
   modules call it). It never consults public_safe_text.

Measured on b15413ffc, calling both owners' matchers directly. A HIT means the text is treated
as private; no row is a HIT in both columns:

value	public_safe_text	public_safety
ghp_… (36 filler chars)	—	HIT
eyJ….….… (JWT shape)	—	HIT
access_key=…	—	HIT
C:\Users\bob\x	—	HIT
the bare word Bearer / password / secret	HIT	—
the Bearer token expired	HIT	—
token=x (short value)	HIT	—
larkoffice…, docs.internal…	HIT	—
Authorization: Basic <base64>	HIT	—
see file:///Users/alice/goal.md	HIT	—
path:/Users/alice/goal.md	HIT	—

The consequence is visible in the code that sits between them: on main,
control_plane/goals/artifact_lifecycle.py writes
if find_private_text_match(value) or _TOKEN_SHAPES.search(value), where _TOKEN_SHAPES is a
private list of provider prefixes carrying this comment: "Provider token shapes the shared
private-text rules do not cover … a leaked token there must never reach a public projection just
because the shared corpus did not list its prefix." That caller knows it needs two owners and
patched the gap locally. #5135 folds _TOKEN_SHAPES and five sibling copies into the
public_safety owner, which removes six copies but still leaves that caller OR-ing
public_safe_text with public_safety, because the prefixes are not the only thing the two corpora
disagree about.

Why this needs an owner decision rather than a merge

The two corpora encode different philosophies, and neither is obviously wrong for all surfaces.
public_safe_text rejects ordinary English words (Bearer, password, secret) — its docstring
records that it deliberately moved one rule away from the bare word because governance prose says
"owner authorization" constantly, while other rules stayed word-shaped. public_safety instead
requires an assignment shape with a length floor so that "the token budget is 1200" survives.
Picking either winner changes what the other set of surfaces is allowed to publish, and it will move
expectations in tests/fixtures/public_safe_text_corpus.json and the TypeScript mirror. There is
also a prefix-blindness dimension to settle: public_safety's path rule excludes a preceding : or
/, which is why file:///Users/alice/goal.md and path:/Users/alice/goal.md pass it while
public_safe_text catches them.

Desired outcome / 期望

One module owns "private-looking text". The other is deleted, or becomes a policy layer that only
adds per-surface thresholds and words on top of the shared shapes. No caller should have to consult
two owners for one value.

Acceptance / 验收

• A single named owner for private-text detection, whose docstring states which rules are
  word-shaped and which are shape-with-threshold, and why.
• artifact_lifecycle no longer ORs two owners for one value.
• The shared corpus fixture pins both directions of the table above, so a future consolidation
  cannot silently drop a rule that only one side had.
• file:///Users/… and path:/Users/… are either rejected by the surviving owner or recorded as
  accepted, with the reason.
• feedback / authority / boundary_authority and validate_public_safe_value are shown to
  reach the same verdicts on a named subset, or each surface's deliberate difference is written
  down next to the owner.

[huangruiteng]

这里有个细节，开发代码提pr时的public safe比loopx内部状态的public safe要严格一些，后者可以松一些