---
name: a-coverage-fix-is-measured-over-the-population-it-claims
description: "A fix that widens a scanner's REACH must be measured over the whole population it claims, not the instance it repaired — and a scan keyed on a structural premise about the codebase must have that premise counted first."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: b2ce30ce-4a9b-4d37-be9e-2469321a0e95
  modified: 2026-08-22T10:02:21.064Z
---

A coverage fix repairs the instance somebody found. Its CLAIM is about a population, and the two
are almost never measured with the same act. Worse, the repair usually installs a NEW structural
premise, and a premise about the shape of a codebase is a countable fact that nobody counts because
it arrives wearing a fix's clothes.

Measured 2026-08-22 at the TURNKEY #212 head gate (19fb90b7). Lane 3's `xtask` spacerun check had
already been caught once by an indented `#[cfg(test)]` latching its scan early — 4 files, 6,375
production lines blind, 5,523 in broker.rs. The fix: only a COLUMN-0 marker ends the scan, on the
stated premise that "a column-0 marker opens the test module that runs to the end of the file."
That premise is false in the largest file in scope. `in_tests` latches and never resets, and cli.rs
carries EIGHT column-0 test modules with production code after each, the first at line 2934 of
37,903. The lane's own `A=0` census had been measured under it, and the requirement text named the
two crate roots with no qualification.

**AND MY OWN HEADLINE FIGURE WAS WRONG IN THE SAME SHAPE, corrected by the builder within the hour.**
I reported "88,030 lines, 54.4%, 102 of 110 files unscanned". That measured LINES AFTER THE FIRST
LATCH and I reported it as COVERAGE LOSS. A file whose only column-0 marker opens its TRAILING test
module lost nothing — that region is out of scope under the old rule and the new one alike. Measured
apart: 88,042 after-latch = 66,637 out-of-scope-either-way + **21,405 production lines actually
hidden, across SIX files, 19,918 of them in cli.rs**. Same day, fourth instance of one shape:
guard vs EXPOSURE · filter vs COVERAGE · a file's tracking vs the CONTENT'S safety · latch position
vs SCOPE. **Name the property you actually measured.** The corrected figure also answered a question
I had queued a mutation for: seven findings in cli.rs and none in the daemon is not a reach that
widened in one file, it is cli.rs being 93% of everything that was ever hidden — arithmetic settled
it and the mutation would have proven nothing either way.

**Why it survived:** the check was GREEN, the fix DID repair the reported instance, and a green from
a scanner is indistinguishable from a green from a scanner that never looked. `A=0` before a
coverage fix and `A=0` after it are the same number for opposite reasons.

**How to apply:** when a fix widens reach, (1) state the structural premise the new rule rests on
and COUNT it across the population — "how many files put a test module mid-file" is one script;
(2) re-run the census over the newly-opened region and expect real findings there; (3) treat any
landed sentence naming the scope as a stater that goes false the instant reach changes, and correct
it by replacement in the same commit. And prove reach the only way it can be proven: a mutation
injected in the region under test with a CONTROL injected where the scanner is known to look —
same class, same run length, same binary, differing only in WHERE. Treatment exit 0 with control
exit 1 is what separates "blind zone" from "inert instrument"; either alone is an argument.
Kin: [[a-mutation-that-reds-every-arm-proves-only-damage]],
[[unapplied-mutation-green-reads-as-a-passing-guard]], [[dont-take-a-diagnosis-as-measured]].