coverage: include/morph/offline is the library's worst subsystem at 85.20%, and one header inside it reads 57.04%

[Yaraslaut]

What is wrong

include/morph/offline is the worst-covered subsystem in the library, and one
file carries almost all of it.

Measured on master (Codecov API), include/morph by subsystem:

coverage	hits/lines	miss	partial	subsystem
85.20%	426/500	56	18	include/morph/offline
86.66%	13/15	2	0	include/morph/qt
94.37%	1158/1227	41	28	include/morph/forms
97.08%	2461/2535	22	52	include/morph/core
97.48%	774/794	2	18	include/morph/util
97.53%	198/203	2	3	include/morph/session
99.25%	266/268	1	1	include/morph/journal
100%	115/115	0	0	include/morph/detail
100%	59/59	0	0	include/morph/render
—	—	—	—	include/morph/net — absent from the report (morph#403)

Library total: 95.69% (5,470/5,716).

Within offline, one file accounts for the great majority of the gap:

include/morph/offline/sqlite_offline_queue.hpp   57.04%   52 misses, 12 partials

52 of the subsystem's 56 misses are in that single header.

The reason to read that number sceptically first

morph_offline_sqlite_tests is a dedicated test executable for exactly this
file, it is coverage-instrumented under AF_COVERAGE, and
scripts/coverage.sh never passes it to llvm-cov (morph#403). So 57.04%
is the coverage this file gets from morph_tests alone, with its own suite's
contribution discarded.

The honest first step is therefore to re-measure with morph#403 fixed, not to
start writing tests. Two outcomes, and they lead different places:

• the number rises sharply — the gap was a reporting defect, and what remains is
  a smaller, real gap;
• the number barely moves — morph_offline_sqlite_tests is thin, which is a
  worse finding than the percentage suggested and belongs in its own right.

What this is aiming at

The goal is include/morph at ≥ 99% lines. From 95.69%, that means closing
about 190 of the 246 currently-unhit lines. offline holds 74 of those 246,
the largest single share of any subsystem, so it is where the target is won or
lost.

Verification status: reproduced

Codecov API for LASTRADA-Software/morph, branch master
(/totals/?branch=master), aggregated per directory. Per-file figures from the
same response.

Not verified: the split between untested and unreachable within those 74
lines. codecov.yml documents a class of llvm-cov closing-brace artifacts that
inflate such counts, and offline's share of them is unknown.

What would close this

1. include/morph/offline reaches ≥ 99% lines, measured after morph#403 —
   or, if it cannot, every remaining line is recorded as unreachable with its
   reason, in the form codecov.yml's existing artifact audits use.
2. sqlite_offline_queue.hpp's figure is re-measured with
   morph_offline_sqlite_tests profiled, and both the before and after
   numbers are recorded, so it is visible how much of the gap was reporting
   and how much was tests.
3. The subsystem carries a codecov.yml component with a target derived from
   its measured ceiling, so the result does not silently regress.

Blocked by morph#403.

[Yaraslaut]

Acceptance reframed: kill named survivors, not reach a percentage

Recording a change of direction rather than editing the text above.

The first framework mutation run (morph#405) scored 64.06% against a library
already measuring 95.69% lines and 91.19% branches. That ~30-point gap is between
executed and would be noticed if it were wrong — and a percentage target is
what produced a suite sitting on the wrong side of it.

So this ticket's coverage-percentage clause is superseded. The work is now driven
by scripts/mutation_survivors.json, which names the missing assertions directly:

• morph#431 — the 255 unasserted survivors in include/morph/core, starting
  with twenty sites where the server's reply(...) can be deleted and the suite
  still passes.
• morph#432 — the 46 in include/morph/forms, mostly boundary comparisons
  exercised only away from their boundaries.

This ticket's non-percentage clauses still stand on their own and are not
superseded — the specific defect it names, the audit it asks for, and any
reachability question it raises are unaffected.

Why the change. The round that produced the measurement added 3013 lines of
tooling under scripts/, cmake/, .github/ and codecov.yml, and 130 lines
under tests/ — of which exactly one was a new test case. Coverage did not
move a point. Sequencing every measurement ticket ahead of every test-writing
ticket is a planning error, and this is the correction: infrastructure is frozen
(morph#430 moved to Backlog) until the survivors are killed.

The guard that comes with it. Do not chase the number. A mutant killed by an
assertion written to kill that mutant, on a path nobody cares about, is worse
than a survivor — it is a test with no reason to exist. Each new test must be
justifiable from the behaviour, with the survivor as evidence the assertion was
missing. Classifying a survivor as equivalent with a stated reason is a valid
outcome; the run already found 11 such.