2026-07-30-sync-coverage-baseline-3.md

Sync — The Coverage Baseline, Third Measurement

Date: 2026-07-30; figure re-taken 2026-08-02 at 048e28a after two session limits interrupted the write-up. The two preserve: commits in this branch's history are where each interruption fell and are left unamended. Status: Measurement. Recorded, reviewed by a human, gated by nothing — the same treatment 2026-07-27-sync-benchmark-gates.md gives every tier B axis, for the reason it gives. Scope: How much of src/sync the default suite executes at 048e28a; which modules are worth hardening next, ranked by what a wrong answer in each one costs rather than by percentage; which of the two earlier baselines' line citations still resolve; and what became of every module either of them named. Supersedes: 2026-07-29-sync-coverage-baseline.md (58257f6) and 2026-07-29-sync-coverage-baseline-2.md (5c546fa). Both remain correct records of the trees they measured and neither is renumbered here, for the reason 2026-07-29-sync-spec-audit-log-2.md gives: a measurement half from one tree and half from another is worse than a dated one. No production code changed. This document is a measurement and a ranked list. It names modules as thin; hardening them is the work it exists to direct, not work it did.

The figure, and how to reproduce it

SYNC_DSN=postgresql://sync:sync@localhost:5433/sync_w118 PYTHONIOENCODING=utf-8 \
  uv run pytest -q --cov=sync --cov-report=term-missing --cov-report=json:coverage.json -rs

97.71% of 7,234 statements, 166 missed, across 97 files — 7,068 covered, 65 further statements excluded outright. Measured on this branch, whose src/ and tests/ are byte-identical to origin/main at 048e28a; the branch adds documentation and nothing else, which git diff 048e28a HEAD --stat shows. The suite reported 2,719 passed and 4 skipped, exit 0, under the -n auto scheduler pyproject.toml's addopts selects, twelve workers on this machine. SYNC_DSN named sync_w118, used by nothing else; under -n auto each worker subdivides that name into sync_w118_gw<n> and creates and drops its own, so the pinned name itself is never created and never dropped.

The figure was taken twice, in two sessions a day apart, and reproduced to the statement: 7,234, 166, 97 files, 65 excluded, both times. Wall clock is not among the reproduced numbers and is not quoted here — the two runs took 168s and 341s, and the difference is machine contention rather than anything about the suite.

-rs is in the command because the skip list is evidence rather than noise here, for the reason the fourth qualification below gives. It changes no figure.

--cov-report=json is the only addition to the command the previous baseline published, and it changes no figure: it writes the same run in the form scripts/rank_coverage.py reads, so the ranking below is computed from the measurement rather than transcribed from it.

Four qualifications travel with the number, and the fourth is new.

The end-to-end test is deselected, as it always is. addopts carries -m 'not e2e', so the figure describes what CI runs.

The denominator is statements, not branches. A line that executes once with one input counts as covered whether or not its other outcome was ever taken, so every figure below is an upper bound on how well a module is exercised. Sixty-five further statements are excluded outright and appear in no denominator here.

Code executed in a subprocess is still invisible to it. Baseline-2 raised this against mcp/server.py, whose main() is driven by tests/test_mcp_entry_point.py through a real pipe and was scored as untested. That module now reads 100% for an unrelated reason — M3-W107 covered the rest of it — so the caveat no longer distorts this table, but the property has not changed and the next module tested through a subprocess will read as thin again.

Four tests skip, and which four depends on the checkout rather than on --cov. An earlier draft of this document claimed that two tests skip under coverage and pass without it, and reconciled a gate run of 2,701 passed and 2 skipped against a coverage run of 2,699 and 4 that way. That claim was wrong, and the way it was wrong is the defect this document exists to find in other modules — a coverage figure taken over a suite the gate does not run. It is recorded rather than deleted because the correction is the useful part.

The two runs were taken in different checkouts. Measured here at 048e28a, in one worktree, back to back, with nothing else changed:

uv run pytest -q -rs                                    : 2,719 passed, 4 skipped
uv run pytest -q --cov=sync ... -rs                     : 2,719 passed, 4 skipped

Identical, to the test. --cov changes no test's outcome on this project. The four skips are the same four in both, and every one of them names a gitignored input:

Skipped	Why	Present where
test_parameter_reduction.py::test_the_reduction_collides_no_stripe_operation[v2320]	.cache/specs/v2320.json absent	fetched, not committed
test_parameter_reduction.py::test_the_reduction_collides_no_stripe_operation[v2330]	.cache/specs/v2330.json absent	fetched, not committed
test_oasdiff_determinism.py::test_repeated_runs_are_nested_rather_than_divergent	opt-in: SYNC_OASDIFF_DETERMINISM=1, and the pinned specifications	never on by default
test_symbol_map_pin.py	.cache/specs/symbols.json absent	fetched, not committed

The two parametrised test_parameter_reduction cases are the whole discrepancy. They pass in any tree where somebody has run scripts/fetch_measurement_inputs.py and skip everywhere else. The coordinator's checkout has .cache/specs/ populated from a fetch on 2026-07-30 — eight Stripe specifications, none of them committed — and this worktree does not. So the same suite at the same commit reports 2,721 passed and 2 skipped there and 2,719 passed and 4 skipped here, and the totals reconcile exactly: 2,721 + 2 = 2,719 + 4 = 2,723.

They always did reconcile, which is what should have prompted the question rather than the arithmetic that was offered instead. The earlier draft was comparing across two commits as well as two checkouts: the coordinator's 2,701 and 2 was measured at d615b75, twenty tests before 048e28a. Those twenty are tests/test_rank_coverage.py and tests/test_replay_shape_writeback.py, which merged in between and collect twenty items exactly. Two variables had moved, and the draft attributed the whole difference to a third that had not moved at all — --cov, which moves nothing.

CI is on this side of it. .github/workflows/ci.yml installs oasdiff and syncs dependencies and never runs scripts/fetch_measurement_inputs.py, so the suite CI gates on skips the same four. The figure above therefore describes the suite CI runs, and it is the populated checkout that is the outlier.

The residual caveat is real but small: those two cases assert that symbols_typescript._comparable is injective over Stripe's 587 operations, and in a checkout without the specifications nothing takes that measurement. It costs no coverage — the reduction is reached by the committed Anthropic and Vercel fixtures either way — but a reader who wants the Stripe half has to fetch first.

What moved, and what a threshold would have made of it

	statements	missed	covered	files	tests
58257f6 — baseline 1	4,916	211	95.71%	84	1,468
5c546fa — baseline 2	5,278	235	95.55%	87	1,569
048e28a — here	7,234	166	97.71%	97	2,719

Between the second measurement and this one the tree grew by 1,956 statements — 37% — and the number of statements nothing executes fell by 69. That is the first movement in this series large enough to mean something, and it is not the percentage: the percentage moved 2.16 points, which is the least interesting number in the table and exactly the one a gate would have been built on.

What produced it is legible in the commit history and is named module by module further down: ten tasks between 5c546fa and here read a single module statement by statement, wrote a decline table for it, and covered what a fixture could reach. Nine of the sixteen modules the two earlier baselines named are now at 100% or one statement short of it.

The ranking, and the weighting it rests on

A percentage ranking answers the wrong question, and both earlier baselines said so in prose beside a table sorted by percentage. This one is sorted by exposure — missed statements times what being wrong in that package costs — computed by scripts/rank_coverage.py, which carries the weights and the argument for each so a reader can disagree with a number rather than with a paragraph.

Tier	Weight	Packages	Why a wrong answer costs this much
patch	4	sync/remediate/, sync/route/, sync/verify/	These choose, write and gate an edit to a customer's source. The worst case this project has is the one CLAUDE.md names — a patch that parses cleanly and means something else — and it is produced here or caught here. verify/ is weighted with them rather than below them because it is the only thing between a wrong edit and a pull request.
signal	3	sync/index/, sync/signals/, sync/detect/, sync/telemetry/	A silent decline here costs a binding, and a run missing a binding reports clean: the vendor broke, the call site was never found, and nothing distinguishes that from a healthy repository. A decline that resolves wrongly is worse and rarer.
record	2	sync/core/, sync/graph/, sync/forge/, sync/benchmark/	These carry what the earlier stages established and publish it. A defect corrupts the record rather than the customer's source — recoverable, but not free: benchmark/ produces the labels precision and recall are gated on.
report	1	sync/cli.py, sync/dashboard/, sync/mcp/	A defect costs a number on a screen or a malformed tool response, with an operator reading it.

Three things about this weighting are worth attacking, and are stated so somebody can.

It is per package, not per statement. A package is the coarsest granularity at which the cost argument is actually true, and the alternative — a weight per module — goes stale the day somebody splits a file. The cost of the coarseness is visible in exactly one row: two of cli.py's uncovered statements (1062–1063) check out a branch and write graph state, which is pipeline wiring and not a number on a screen. Weighted at 1 with the rest of the entry point, they are under-counted.

sync/cli.py is deliberately the low weight even though it is where the pipeline is wired. The wiring failure this project actually keeps having — a component finished, tested, and reachable from nothing — is invisible to line coverage in both directions, and scripts/lint_dead_links.py is the gate for it. Weighting cli.py up would be using this instrument to chase a defect it cannot see.

sync/mcp/ at 1 is arguable. It is the product surface a customer's agent talks to, so a wrong answer there is a wrong answer delivered to a model rather than to a person. It sits at 1 because a malformed tool response is observed by the agent that receives it, and 2026-07-30-provenance-on-the-tool-surface.md is the case for reading that as recoverable. Anyone who disagrees should raise it to 2 and re-run; at one missed statement in the whole package it changes no ordering today.

Every module with something uncovered, ranked

Generated by uv run python scripts/rank_coverage.py coverage.json. The last column is added by hand and is the one that matters most: examined means a committed report has already read those statements one at a time and classified each.

Module	Tier	Missed	Statements	Covered	Exposure	Already examined?
sync/index/python_lang.py	signal	25	373	93%	75	No
sync/cli.py	report	31	573	95%	31	No
sync/index/typescript.py	signal	10	307	97%	30	No
sync/remediate/nodes.py	patch	7	213	97%	28	No
sync/route/templates.py	patch	6	229	97%	24	Yes — 2026-07-30-edit-primitive-declines.md, all six
sync/signals/generated/symbols_typescript.py	signal	8	274	97%	24	Yes — 2026-07-29-typescript-symbol-reader.md, all eight
sync/signals/registry.py	signal	7	157	96%	21	No
sync/core/conformance.py	record	9	287	97%	18	No
sync/benchmark/mutate.py	record	8	213	96%	16	Yes — 2026-07-29-mutation-engine-declines.md, all eight proven unreachable
sync/signals/generated/symbols_speakeasy.py	signal	5	245	98%	15	Yes — 2026-07-30-speakeasy-reader-declines.md, the same five lines
sync/signals/oasdiff.py	signal	5	69	93%	15	Partly — four are the correctly-thin binary lookup both earlier baselines named
sync/remediate/parameters.py	patch	3	58	95%	12	No
sync/signals/generated/adapter.py	signal	4	142	97%	12	No
sync/signals/sentry/shapes.py	signal	4	83	95%	12	No
sync/detect/status_rate.py	signal	3	119	97%	9	Yes — 2026-07-29-detector-declines.md, 2026-07-30-declines-that-cost-findings.md
sync/signals/datadog/shapes.py	signal	3	87	97%	9	No
sync/signals/deprecations/adapter.py	signal	3	74	96%	9	No
sync/signals/deprecations/catalogue.py	signal	3	172	98%	9	No
sync/detect/observed_drift.py	signal	2	89	98%	6	Yes — 2026-07-29-detector-declines.md
sync/index/dependency_edits.py	signal	2	72	97%	6	No
sync/index/tsc.py	signal	2	59	97%	6	No
sync/forge/github.py	record	2	160	99%	4	No
sync/remediate/agent_patch.py	patch	1	69	99%	4	No
sync/remediate/corpus.py	patch	1	82	99%	4	No
sync/remediate/tiered.py	patch	1	98	99%	4	No
sync/verify/mock_response.py	patch	1	74	99%	4	No
sync/verify/replay.py	patch	1	104	99%	4	Partly — the timeout, named correctly thin twice
sync/index/literals.py	signal	1	52	98%	3	Yes — 2026-07-29-python-flavour-and-literals.md
sync/index/shipped_tree.py	signal	1	55	98%	3	No
sync/signals/generated/manifest.py	signal	1	71	99%	3	No
sync/signals/generated/symbols.py	signal	1	176	99%	3	Yes — 2026-07-29-python-flavour-and-literals.md
sync/benchmark/score.py	record	1	80	99%	2	No
sync/core/models.py	record	1	196	99%	2	No
sync/forge/webhook.py	record	1	54	98%	2	No
sync/dashboard/queries.py	report	1	78	99%	1	No
sync/mcp/tools.py	report	1	93	99%	1	Yes — 2026-07-30-mcp-tool-surface-declines.md

Sixty-one further modules are at 100% and are not ranked, because a ranking of what to harden has nothing to say about a module with nothing uncovered.

The ordering is stable against the one filter that could have changed it: the top four rows are the same whether or not the already-examined modules are removed first. route/templates.py and symbols_typescript.py would otherwise sit fifth and sixth, and both have been read statement by statement already.

What to harden next: three modules, and the reason each

src/sync/index/python_lang.py — 25 missed, exposure 75, the largest by a factor of two.

It has held that position in all three measurements, and it is the only module that has. Baseline-1 named it and one repair was made — tests/fixtures/py/aliased and tests/test_python_aliases.py, recorded in the audit log — which closed the specific asymmetry that document argued from and left the module thin. Its statement count has grown from 278 to 373 since then and its uncovered count from 23 to 25.

Eighteen of the twenty-five are literal declines: return None, return False, continue. That is the shape every decline report on this project has found, and the same instrument applies. The other seven are not, and two of them are capability rather than guard — line 319–320 renders the count of declared requirements, and line 867 answers "mypy" for a repository that configures it through a file rather than through pyproject.toml.

src/sync/index/typescript.py — 10 missed, exposure 30 — and it should be hardened in the same task, because one fixture closes a statement in both.

python_lang.py:628 and typescript.py:495 are the same statement in the two indexers: _enclosing_scope walking to the root and returning it. Both docstrings say what reaches it — the TypeScript one in as many words, "falls back to root for a module-level call, which has no such ancestor" — and neither is covered, which means no fixture in this repository binds a call result at module scope in either language. Every response-side fixture wraps its call in a function. A customer's script that calls stripe.charges.create at the top of a file and reads a field off the result is ordinary, and nothing has ever shown either adapter handles it.

That is the same species of finding as the alias asymmetry baseline-1 opened with, and it is worth saying that it is worse: that one was an asymmetry between the two adapters, visible to anyone comparing the two fixture directories. This one is symmetric, so comparing the adapters finds nothing, and only the coverage number points at it.

src/sync/remediate/nodes.py — 7 missed, exposure 28, and the highest cost per statement in the table.

Named third rather than cli.py, and the ranking is why: cli.py outranks it on volume alone at the lowest weight, while these seven sit on the patch path at the highest. Five of the seven are two except Exception handlers — the ones that turn a fault inside the patch cascade into a recorded decline instead of an abandoned run — and neither has ever caught anything in a test. CLAUDE.md's rule about abandoned runs being data rests on exactly those two handlers writing an abandon_reason a router can learn from, and nothing has shown either one writes it.

The remaining two (342, 347) are the message rendered when a patched call path fails replay, which is the sentence a human reads on a declined finding.

sync/cli.py is second by exposure and is deliberately not named here. Thirty-one statements is the largest raw gap in the tree and has been in all three measurements, but the module is the command line: much of the gap is argument handling, json.load(sys.stdin), and one whole key-printing subcommand at 1503–1512. It is worth a task; it is not worth the next task, and this is the disagreement the weighting exists to make explicit rather than to hide.

Where coverage has nothing left to tell you, and what would

Forty-eight per cent of the missed statements in this tree are literal declines — 79 of 166, or 47.6%. The count is every missed line whose stripped text is exactly one of continue, pass, return None, return False, return [], return {}, return () or return set(); that list is given in full because a proportion computed from an unstated set of forms is not a measurement anybody can check. In the two indexers at the top of the ranking the proportion is higher: 18 of 25 and 8 of 10.

For a module in that shape, the coverage number has already told you everything it can. It says a return None never ran. It cannot say whether the return None is right, what input reaches it, or what the caller sees when it fires — and this project has now measured, ten times, that the answer to the last question is usually nothing at all. 2026-07-29-hand-written-symbol-maps.md found eleven declines and eleven times "the caller observes nothing"; 2026-07-30-declines-that-cost-findings.md found eight that cost a finding and one that costs nothing, which is the first time the answer differed.

Worse, a covered decline is not a tested one. A continue that some distant test perturbs for an unrelated reason reads as covered and is checked by nothing, and the ten reports above found this repeatedly — 2026-07-29-hand-written-symbol-maps.md's row 6 is a guard that was uncovered only because a fixture reduction script deleted its input, and the same file names a covered line carrying a sub-condition that can never be false.

So for these modules the instrument to reach for is not more coverage. It is the one this project has converged on independently in ten tasks and which docs/superpowers/reports/2026-07-29-hand-written-symbol-maps.md gives the canonical shape of:

• A decline table. One row per uncovered statement: the statement, the input that reaches it, whether declining is the right answer for that input, and what the caller observes. The fourth column is the one that produces findings; three separate reports found a decline no downstream artifact could count.
• A mutation pass over the statements the table covers, to establish that the tests added are not vacuous. Every one of these reports pins existing behaviour, so no test in them was red before the code existed, and mutation is what stands in for that.
• A named fixture for the input, where one exists. python_lang.py:628 and typescript.py:495 need a module-scope call site, and that is a fixture, not a redesign — the same answer baseline-1 reached about the alias gap.

sync/benchmark/mutate.py is the clearest case of coverage having stopped being informative. It sits ninth in the ranking at eight missed statements, and all eight were proven unreachable by three independent kinds of evidence — structural, empirical over 372 committed source files, and a whole-suite run with all eight replaced by raise AssertionError. Nothing will ever cover them and nobody should try. The ranking cannot know that, which is why the last column of the table exists.

The previous baselines' line citations, and which still resolve

The spec audit named one broken citation and declined to renumber it. There are nine in the two documents. Eight no longer resolve; one does.

Citation	Document	What it described	What sits there at 048e28a	Where the construct lives now
index/python_lang.py:206-210	baseline 1	from stripe import X as Y	the vendor-binding constructor reading distribution	437–443
index/python_lang.py:220-223	baseline 1	import stripe as s	a comment and self._symbol_root	453–456
index/python_lang.py:308-309	baseline 1	walking a nested dictionary argument	an error-message f-string	_dictionary_paths, 562–585, called from 609–612
mcp/server.py:88	baseline 2	the continue on a blank line in the stream	for line in source:	90
mcp/server.py:200-201	baseline 2	the except Exception tool-fault guard	return _result(	225
signals/oasdiff.py:26-29	baseline 2	the binary not being found	a comment on exit levels	49–52
signals/oasdiff.py:72	baseline 2	a non-zero exit from the binary	_parse_json(result.stdout, ...)	70–71
verify/replay.py:221	baseline 2	node absent from PATH	if node is None:	221 — resolves
verify/replay.py:344-345	baseline 2	the sixty-second timeout	env=_environment(...)	350–351

One correction to the audit log, which is a fact about how fast this drifts rather than a fault in the audit. Its replacement reading — that python_lang.py:206-210 "is now a set comprehension inside _bindings_for" — was true at 0613da2 and is not true here: _bindings_for no longer exists anywhere in the file. A citation and its correction both went stale inside one day.

This is the argument for a successor rather than an edit, and it is stronger than the audit could state it. Eight of nine citations moved: baseline-1's three over the 440 commits from 58257f6 to 048e28a, and five of baseline-2's six over the 416 from 5c546fa. Renumbering them would have produced a document whose tables were measured at one commit and whose line ranges pointed at another, which is the failure mode the audit refused and was right to refuse.

Every module the two earlier baselines named, and what happened to it

Sixteen modules were named across the two documents. None is gone; every file still exists. Nine are at 100% or one statement short of it, two have had every remaining statement classified by a committed report, and five still carry uncovered statements nobody has read.

Module	58257f6	5c546fa	048e28a	Fate
telemetry/otlp.py	83%, 12 missed	83%, 12 missed	100%, 0	Repaired by baseline-2's own task, thirteen tests, each proven non-vacuous by mutation
mcp/server.py	86%, 13	82%, 17	100%, 0	Repaired — 2026-07-30-mcp-tool-surface-declines.md (M3-W107)
signals/mcp_server/arguments.py	89%, 4	89%, 4	100%, 0	Repaired — 2026-07-29-mcp-signal-refusals.md
signals/mcp_server/snapshot.py	93%, 2	—	100%, 0	Repaired — same report
signals/twilio/symbols.py	90%, 6	90%, 6	100%, 0	Repaired — 2026-07-29-hand-written-symbol-maps.md (M3-W95)
signals/stripe/symbols.py	92%, 5	92%, 5	100%, 0	Repaired — same report
verify/mock_response.py	89%, 8	89%, 8	99%, 1	Repaired but for one statement
verify/replay.py	90%, 10	90%, 10	99%, 1	Repaired but for the sixty-second timeout, which both earlier documents called correctly thin
index/literals.py	92%, 4	92%, 4	98%, 1	Repaired — 2026-07-29-python-flavour-and-literals.md; the remaining statement is that report's
benchmark/mutate.py	—	92%, 12	96%, 8	Closed. All eight remaining proven unreachable — 2026-07-29-mutation-engine-declines.md
route/templates.py	95%, 12	95%, 12	97%, 6	Closed. All six examined and classified — 2026-07-30-edit-primitive-declines.md
index/typescript.py	94%, 16	93%, 18	97%, 10	Improved; still thin, third in the ranking, and shares a statement with the module below
index/python_lang.py	88%, 33	92%, 23	93%, 25	Still thin, and first in the ranking for the third time. One repair made (the alias fixture); the rest never taken up
core/conformance.py	—	88%, 5	97%, 9	Grew from 43 to 287 statements and stayed roughly as thin proportionally. Still unexamined
signals/oasdiff.py	92%, 5	92%, 5	93%, 5	Unchanged in three measurements. Four of the five are the binary lookup both baselines called correctly thin; the fifth is a non-zero exit from oasdiff checks, which a fixture reaches
cli.py	95%, 18	94%, 27	95%, 31	Grew with the tree. Largest raw gap in all three measurements and never acted on

So the recommendations were tracked, and this is the first of the three baselines that can say so. Baseline-1 named one module to harden and its repair is recorded in the audit log. Baseline-2 named one and hardened it in the same task. The eleven repairs above were mostly not made because a baseline asked for them — they were made by tasks reading one module at a time — but the overlap is near total, and the two modules the baselines named that nobody took up are exactly the two still at the top of this ranking.

That is the honest reading, and it has one uncomfortable half. index/python_lang.py has been named first in three consecutive measurements over 440 commits and has been hardened once, in a way that closed the specific example the document argued from rather than the module. A baseline whose top recommendation survives three of its own successors is a document being read and not acted on.

The --cov trap, and whether this measurement hit it

No measurement in this document used a dotted --cov argument. Every figure comes from --cov=sync, and there are no per-module runs.

That matters because a per-module coverage run on this project can fail for a reason that has nothing to do with the module. 2026-07-29-psycopg-error-identity.md established the mechanism, and it survived one retraction that was itself wrong, so it was re-verified here at 048e28a against the installed coverage 7.15.2 rather than cited:

--cov=sync                   : 0 psycopg modules imported inside sys_modules_saved(), 0 evicted
--cov=sync.benchmark.mutate  : 77 imported, 77 evicted on restore
                               psycopg.Error is the same class          = False
                               isinstance(first-set exc, psycopg.Error) = False
                               isinstance(first-set exc, first Error)   = True

Under a second, no pytest involved, against coverage 7.15.2 and psycopg 3.3.4.

The probe has a failure mode that reports the trap as absent, and it was hit while re-verifying this. A probe that imports psycopg before entering sys_modules_saved() measures zero modules imported and zero evicted for both arguments, and every identity check then answers True — because the modules are already in sys.modules, so they are not new to the saver and restore() deletes none of them. That reads exactly like a fixed bug. The probe must reach file_and_path_for_module with psycopg unloaded, which is what the real sequence does: coverage resolves its source argument at startup, before any test has imported anything. Anyone re-checking this should assert "psycopg" not in sys.modules at the top of the probe, and treat an all-True result as a broken probe until that assertion has passed.

coverage.inorout resolves a dotted source argument by importing it inside coverage.misc.sys_modules_saved(), whose restore() deletes every module the import added; psycopg/errors.py is pure Python and re-executes into a second set of exception classes, while the psycopg_binary._psycopg extension is cached below sys.modules, is not re-executed, and goes on raising the first set. --cov=sync names a package that imports no psycopg, so nothing is evicted and nothing splits.

Anyone taking a per-module figure to check one of the modules named above should expect this and say whether their run was clean. A dotted --cov is exactly the natural way to measure a single module, which is what makes the trap expensive: it fires on the measurement, not on the code, and the resulting red test is attributed to whatever was being measured.

What this number still cannot see

Unchanged from both earlier baselines, and restated because it has not stopped being true.

Line coverage cannot see whether anything in production reaches the code at all. Seven components shipped here fully covered and reachable from nothing, and a line count called every one of them healthy; an eighth arrived in the commit that added it. scripts/lint_dead_links.py is the gate that catches that class, and the number above is not.

Nor can it see whether an assertion is any good. Ten decline reports on this project reached the same conclusion from ten different modules: the statements a coverage number points at are the cheap half of the question, and the expensive half — what the caller observes when the decline fires — is invisible to it in both directions.

So the figure is worth what it is worth. It says the suite executes most lines of most modules. It does not say the assertions are good, that the branches not taken behave, or that a line of it is reachable from an entry point. A 97%-plus figure arriving without that sentence will be trusted further than it deserves.

Why no threshold, a third time

2026-07-27-sync-benchmark-gates.md binds: "do not invent a threshold. A gate at an invented number either fires constantly and gets disabled, or never fires and provides false assurance." CI records the figure with || true and gates on nothing, which is correct and was left alone.

There are now three measurements, at 58257f6, 5c546fa and 048e28a. Three points are the beginning of a history rather than a history, and this measurement finally moved by more than the noise the second one had to account for — 2.16 points, against a run-to-run variation baseline-2 measured at three statements. So the objection is no longer that the signal is smaller than the noise.

It is now the other objection, and it is the stronger one. The exposure column in this document is a ranking key, and turning a ranking key into a floor is how it stops measuring. A percentage that fails a build is a percentage people write tests to satisfy, and this repository's rule about a test that cannot fail applies exactly as well to a test written to move a number. The gate tier C describes as safe — a directional floor on a deterministic measurement — is still the only one that would be safe here, and what it would floor is the count of unexamined uncovered statements, not the percentage. That is a number a decline table moves and a vacuous test does not, and it is worth building only once somebody has written the fourth of these documents and can say whether the column above went up or down.