ci: split mutation_survivors.json's citations by what they are, and fold the NOLINT gate into drift-guard (fixes #613, fixes #633)

.github/workflows/drift-guard.yml

@@ -3,8 +3,13 @@ name: Drift guard
 # Gates for the "a list that stopped matching what it describes" class of bug.
 # Every job here is fast and dependency-free; none of them compiles anything.
 #
-# Sections below: the spec <-> code gates, the ladder rung-list gate, then the
-# scenario-coverage gate.
+# It also carries one gate for the neighbouring "a suppression that suppresses
+# nothing" class (the NOLINT-directive scan at the end of prose-lint, #627),
+# which is the same shape of defect one level down: a control that parses, looks
+# applied, and measures nothing.
+#
+# Sections below: the spec <-> code gates and the NOLINT-directive gate, the
+# ladder rung-list gate, then the scenario-coverage gate.
 #
 # Two independent gates for the "spec <-> code drift" class of bug (a
 # docs/spec/*.md file stating a mechanical fact -- an enum cardinality, a
@@ -157,6 +162,46 @@ jobs:
       - name: Check every test name selects itself as a Catch2 filter
         run: bash scripts/check_catch_test_names.sh
 
+      # ── NOLINT directives that cannot take effect (#627, folded here by #633) ──
+      # NOLINTNEXTLINE applies to the next *physical* line. When the reason is
+      # wrapped onto a second comment line, the directive annotates that
+      # comment, the findings it names are still reported, and clang-tidy says
+      # nothing about it -- the directive parses and the file looks annotated.
+      # Four such directives existed on master and six findings leaked past
+      # three of them (#627).
+      #
+      # That is the failure mode AGENTS.md names first -- a control that
+      # reports success while measuring nothing -- one level below the gate. It
+      # also matters for the clang-tidy campaign in #580, whose agreed remedy
+      # is partly "write an individually reasoned NOLINT": that campaign cannot
+      # proceed honestly on a tree where NOLINTs silently fail to apply, so
+      # this gate is its precondition.
+      #
+      # A fast, dependency-free text scan over ~700 source files that compiles
+      # nothing, so it rides this job for the same reason the CI-clang-pin and
+      # Catch2-name scans above do, rather than paying for a runner of its own.
+      # It landed as a workflow of its own only because this file was held by
+      # an open PR at the time (#614); that PR has landed and #633 folds it in.
+      #
+      # The scan leaves a deliberate residual: an anchoring `NOLINTNEXTLINE`
+      # whose own line carries no finding is accepted, because a directive may
+      # legitimately annotate a line the current check set says nothing about.
+      # The self-test below pins that, and it is the reason this gate is not a
+      # blanket "every directive must suppress something".
+      #
+      # Self-test first, as every other lint in this repository does: a gate
+      # that detects nothing reports the same green as a clean tree, and for
+      # this gate that would be the very defect it exists to catch, one level
+      # up. It drives the checker against tests/lint/nolint_directives/ -- the
+      # two effective shapes must be accepted, each inert shape rejected on its
+      # own and naming its own file, and a directory with no directives at all
+      # rejected rather than called clean.
+      - name: Self-test the NOLINT-directive checker
+        run: bash scripts/test_check_nolint_directives.sh
+
+      - name: Check every NOLINTNEXTLINE annotates a line of code
+        run: bash scripts/check_nolint_directives.sh
+
   # ── Ladder rung list <-> the CI filters that are supposed to track it ──
   # examples/rungs.txt is the ladder's single authoritative rung list, and
   # almost every consumer now derives from it at run time. Three cannot --

scripts/check_mutation_survivors.py

@@ -46,15 +46,29 @@
 ----------------------------------------------------------------
 Only the *structured* entries -- dicts carrying `file`, `line` and `source`.
 The same document also carries citations as free text, inside prose strings:
-`"include/morph/core/backend.hpp:613 -- emitMetric(Metric::registerCount, 1.0)"`,
-`"site": "remote.hpp:1031"`, and 29 more (31 in all, as this script's own
+`"site": "remote.hpp:1031"`, `"core/registry.hpp:54 -- PairKeyHash
+hash-combine"`, and 18 more (20 in all, as this script's own
 count_free_text_citations() reports on every run). Those have no verbatim
 `source` text to resolve against -- the text after the `--` is a paraphrase,
 and several use a bare filename with no directory -- so this resolver cannot
-see them, and several of them are measurably stale (morph#613). The count is
-printed on every run so this gate's coverage of the file is visible rather than
-assumed: auditing the structured half of a file while the prose half rots is a
-narrower claim than a green tick looks.
+see them.
+
+Every one of those 20 is *deliberately* out of reach, and morph#613 settled
+why. They live in the file's dated campaign records -- `runs`,
+`false_positive_finding`, `mechanism_confirmed` and
+`classification_2026_09_09` -- whose line numbers describe the tree as it was
+when the campaign ran, and each of those sections now carries a `revision` so
+the citation resolves with `git show <revision>:<path>`. Auditing them against
+HEAD would fail forever and correctly so: the record is not wrong, the reading
+would be. The citations that *do* describe current code all sit under
+`classes`, and morph#613 converted them to the structured shape above, which
+this script picked up with no edit to it -- find_entries() walks the whole
+document by design, exactly so that a new class of live citation is covered
+without a second mechanism.
+
+The count is printed on every run so this gate's coverage of the file is
+visible rather than assumed: auditing the structured half of a file while the
+prose half is exempt by design is a narrower claim than a green tick looks.
 
 Vacuity
 -------