fix: the chip drift guard was watching a dict nothing used

[quadseven]

Size: XS

Why

_EFFORT_LABELS was defined and never read. Its test asserted its keys equalled
EFFORTS - a guard on a derived dict, which by construction cannot drift,
so the test could never fail.

Meanwhile _SEVERITY_CHIP and _EFFORT_CHIP - the dicts actually rendered into
every marking - are hand-written and had no guard at all. Adding a level to
review_types would fall through .get() to the bare identifier with nothing
to say so.

Which is precisely the failure the old comment claimed to prevent:

Derived from the shared vocabulary so a new effort level can never silently
drop its chip

The guard was watching a decoy.

What changed

• Deleted the dead _EFFORT_LABELS
• Added _chip_vocabulary_drift(), checking SEVERITIES and EFFORTS against
  the dicts that are actually rendered
• Returned rather than asserted at import: a drifted table should degrade to an
  ugly chip in production, not refuse to import dispatch. The test is where it
  fails loudly
• The test removes a level, asserts the guard goes red, then restores it. A
  guard nobody has watched fail is a guess

Acceptance criteria

• The guard covers both dicts that are actually rendered
• The guard is demonstrated failing, not just passing
• A drifted chip table does not break module import
• The dead dict and its tautological test are gone

Out of scope

• The emoji in the chips themselves. They are non-ASCII, but they are a visible
  product choice in the inline markings, not something to change in passing.
• Whether Effort should gain more levels. This only ensures that if it does,
  the omission is loud.

Verification

• 1836 passed, 2 skipped