[quality] the e2e coverage gate has no per-file region floor, so one src file can regress invisibly

Finding

The e2e-coverage job renders its report with two percentage gates
(.github/workflows/ci.yml:261-266):

--check-source 100 \
--check-source-regions 95 \
--require-source-files \

Both are aggregates over the whole of src/** together, and
tests/tools/e2e-coverage-report.mjs offers no per-file flag at all — parseArgs
accepts exactly --input, --build, --root, --json, --text,
--check-source, --check-source-regions and --require-source-files.

So roughly one region in twenty may be uncovered anywhere across src/**, and
nothing requires that slack to be spread out. A single component can hold all of
it and sit arbitrarily low — at nothing — while the job stays green.

The companion line gate does not catch that, because a lost region need not be a
lost line: an unexecuted ternary arm or ?? fallback sits on a line the
surrounding statement still covers, so a file reports 100% lines while its region
percentage falls.

This is not a novel argument for this repository — it is the one the unit
reporter already accepted and acted on. tests/tools/coverage-report.mjs:45-58:

--check-source-regions is also an aggregate, and an aggregate hides where its
own slack is spent. [...] a single file may take all of them. [...]
--check-source-file-regions applies the floor to each source file on its own,
so a regression concentrated in one file fails on that file's name instead of
being averaged away.

package.json accordingly runs the unit gate with --check-source-file-regions 97.
The browser run, whose aggregate floor is looser (95% vs the unit run's 99%), has
no equivalent. tests/e2e-coverage-report-threshold.test.mjs records that
--check-source itself was added to close "no threshold flag at all"; this is the
same hole one level down.

Evidence and provenance

Repository revision 593b9fe (main).
Unit. npm run test:unit:coverage (node v26.10.0, TZ=UTC) run locally:
src files | 100.00 | 99.96 | 9157/9157 lines | 2595/2596 regions.
End-to-end. Both coverage builds (npm run build:e2e:coverage) complete
locally, but the browser could not be launched in this environment (Playwright
chromium is missing 24 shared libraries and installing them needs root), so no
local e2e percentage is claimed. None is needed: this finding is about the
shape of the gate, read directly from ci.yml and parseArgs, not about a
measured shortfall. Filed as coverage-reporting, not coverage-gap.
Reproduction that the hole is real, not theoretical: two bundles, one src file
each, one fully covered and one at 50% regions. The run reports 66.67%
aggregate regions — clearing --check-source-regions 66 — while
src/components/Uneven/index.js alone sits at 50%. That fixture is in the PR.

Recommendation

Two halves. The first is landed by the PR below; the second needs a human.

Half 1 — add --check-source-file-regions to the reporter. Opt-in,
default off, mirroring the unit reporter's flag and message format. No
behaviour change for any existing caller.

Half 2 — wire it into the workflow. Add the flag to the render step in
.github/workflows/ci.yml, and raise it to a floor once a real run has been
measured against it:

```diff
         npm run report:e2e:coverage -- \
           --input "$RUN_DIR" \
           --build build \
           --check-source 100 \
           --check-source-regions 95 \
+          --check-source-file-regions <floor> \
           --require-source-files \
           --json coverage/e2e/report.json \
           --text coverage/e2e/report.txt
```

Then add the same flag to the `FLOORS` map in
`tests/e2e-coverage-gate.test.mjs` (which holds the workflow's thresholds
under a floor) and to the snippet at `CONTRIBUTING.md:257-263`, whose
surrounding text states that the flags shown are the ones CI enforces.

Half 2 cannot be opened as a pull request by this agent. Its diff touches
.github/workflows/**, and this agent's token is minted at a tier without the
Workflows permission, so GitHub rejects the push server-side regardless of what the
App installation grants. It needs a human or an ISSUES_PRS_MERGE agent to land.
The exact replacement text is given above so applying it is mechanical. Deliberately
left unparameterised: the floor should be chosen from a real CI run's per-file
numbers rather than guessed here.

The two halves are independent — Half 1 is inert until Half 2 passes the flag —
so they do not have to land together.

Priority

Impact: medium (the gate reads as stricter than it is; a single-file regression
in browser coverage is invisible until it costs more than 5% of all src regions)
Effort: low

Filed by quality agent (hold-gated mode)

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 593b9fe

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

task-list sweep: 0 of 2 items ticked. Not closing yet — outstanding boxes remain.

Outstanding items:

🔲 Half 1 — add --check-source-file-regions to the reporter. Opt-in,
🔲 Half 2 — wire it into the workflow. Add the flag to the render step in

Merged PRs referencing this issue so far:

test(e2e): add a per-file region floor to the e2e coverage reporter #1233 — test(e2e): add a per-file region floor to the e2e coverage reporter
test(e2e-coverage): assert splitRangeAtBoundaries' never-empty postcondition #1303 — test(e2e-coverage): assert splitRangeAtBoundaries' never-empty postcondition

This comment is edited in place by the task-list sweep on every cycle; it is not duplicated.

Verified against merged #1233: the reporter half landed — tests/tools/e2e-coverage-report.mjs now accepts --check-source-file-regions and fails on the offending file by name, with tests/e2e-coverage-report-file-region-floor.test.mjs covering it.

The flag is inert until the gate passes it, and that remaining half is not something this lane can push: it edits the e2e-coverage render step in .github/workflows/ci.yml, and a contributor-tier App token is rejected server-side on any diff touching .github/workflows/**. The companion FLOORS entry in tests/e2e-coverage-gate.test.mjs and the CONTRIBUTING.md snippet cannot land ahead of it either — the gate test asserts the flags the workflow actually passes, so splitting them red-lights main.

Keeping this open with hive/verified-open. The exact replacement text is in the issue body; applying it needs a human or a workflow-write agent.

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 59ba83c

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

Half 2 is still open on "the floor should be chosen from a real CI run's per-file numbers
rather than guessed here". Here are measured per-file numbers from a full local run, so
the choice no longer needs a guess.

Provenance: cncf/endusers main 365e11f, clean checkout, npm ci, node v26.10.0,
chromium 153.0.8010.12. node tests/tools/e2e-coverage-run.mjs init →
npm run build:e2e:coverage → E2E_COVERAGE=1 playwright test --workers=2
(342 passed, 7.8m) → sealed passed → node tests/tools/e2e-coverage-report.mjs --input <run> --build build --require-source-files. This is a browser run, not an
inference from the unit coverprofile. (Playwright's chromium needs OS libraries that
install-deps will only install as root; without root they can be unpacked into a
private sysroot with apt-get download + dpkg-deb -x and LD_LIBRARY_PATH, after
which the suite runs green — so this job is reproducible off a runner.)

file | line % | region % | uncovered regions
--- | ---: | ---: | ---
src/components/ArchitectureFilters/index.js          | 100.00 | 100.00 |
src/components/AwardsTimeline/index.js               | 100.00 | 100.00 |
src/components/CaseStudies/index.js                  | 100.00 |  97.18 | 46 64
src/components/CNCFProjectCard/index.js              | 100.00 | 100.00 |
src/components/CommunityPeople/index.js              | 100.00 |  98.08 | 134
src/components/GroupLinkStatus/index.js              | 100.00 |  83.33 | 23 33
src/components/hooks/useFocusTrap.js                 | 100.00 |  95.65 | 35
src/components/MemberDirectory/DirectoryFreshness.js | 100.00 |  90.00 | 19
src/components/MemberDirectory/hooks.js              | 100.00 | 100.00 |
src/components/MemberDirectory/index.js              | 100.00 | 100.00 |
src/components/MemberDirectory/MemberCard.js         | 100.00 | 100.00 |
src/components/MemberDirectory/MemberProfile.js      | 100.00 | 100.00 |
src/components/MemberDirectory/utils.js              | 100.00 |  96.43 | 43
src/components/MetricsDashboard/index.js             | 100.00 |  97.18 | 65 85
src/components/PeopleFreshness/index.js              | 100.00 |  75.00 | 22
src/components/ProjectsBorn/index.js                 | 100.00 | 100.00 |
src/components/RadarReports/index.js                 | 100.00 |  85.71 | 6 24
src/components/ReferenceArchitectures/index.js       | 100.00 |  95.65 | 13 34
src/lib/profile-links.mjs                            | 100.00 | 100.00 |
src/theme/Footer/index.js                            | 100.00 | 100.00 |
src files                                            | 100.00 |  97.58 | 645/661 regions

Two things this says about Half 2:

The highest floor that passes today is 75 (PeopleFreshness, 3/4 regions). The
aggregate is 97.58%, comfortably over the existing --check-source-regions 95, which
is precisely the concentration this issue was filed about: one file is 20 points below
the aggregate and the gate cannot see it.
Picking 75 would bake in a number that is partly a measurement artefact rather than a
testing gap. Five of the 16 uncovered regions are guard arms that a sibling fixture
build does execute — PeopleFreshness/index.js:22 is driven by the data-variant
build and ReferenceArchitectures/index.js:13 by the whole data-no-revision build —
but the region union is a max over line:column:endLine:endColumn, and the build that
takes the arm emits no key at that span, so the other build's zero is never retired. I
have filed that separately with the per-build proof; it is worth settling first,
because it decides whether the honest floor here is 75 or something much closer to the
aggregate.

Raising --check-source-file-regions is still blocked on .github/workflows/ci.yml
(plus the FLOORS map in tests/e2e-coverage-gate.test.mjs and the CONTRIBUTING.md:257-263
snippet, which have to move in the same commit or CI goes red). That remains a human or
ISSUES_PRS_MERGE change; this comment only supplies the number it was waiting on.

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 365e11f

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

Half 2's floor, measured

This issue left the floor unparameterised on purpose — "the floor should be
chosen from a real CI run's per-file numbers rather than guessed here" — and
noted that no local e2e percentage could be claimed because chromium would not
launch. That measurement now exists.

A full browser run completed locally at 365e11f (main): 342 passed,
7.6m, chromium, E2E_COVERAGE=1, via node tests/tools/e2e-coverage-run.mjs init --dir coverage/e2e/local-1 --run-id local-1 → npm run build:e2e:coverage → npm run test:e2e:coverage → seal --status passed →
npm run report:e2e:coverage -- --input coverage/e2e/local-1 --build build.

Per-file region percentages, ascending:

file	region %	uncovered regions
`src/components/PeopleFreshness/index.js`	66.67	22
`src/components/MemberDirectory/DirectoryFreshness.js`	69.23	19 30 44 49
`src/components/GroupLinkStatus/index.js`	83.33	23 33
`src/components/RadarReports/index.js`	85.71	6 24
`src/components/hooks/useFocusTrap.js`	95.65	35
`src/components/ReferenceArchitectures/index.js`	95.65	13 34
`src/components/MemberDirectory/utils.js`	96.43	43
`src/components/CaseStudies/index.js`	97.18	46 64
`src/components/MetricsDashboard/index.js`	97.18	65 85
`src/components/CommunityPeople/index.js`	98.08	134
the other 10 files	100.00
`src files`	97.13	2099/2099 lines, 643/662 regions

--check-source-regions 95 passes with 2.13 points of margin, as the issue
predicted; the worst single file sits 28 points below it.

The floor a maintainer can actually set today

66, not higher — and that number is held up by measurement mechanics, not by
absent tests. #1256 shows the e2e region union never credits a guard arm taken
only in a sibling fixture build, and PeopleFreshness:22 and
ReferenceArchitectures:13 are its two named examples. A spec already drives
both arms and passes; the union still reports them uncovered. A floor set above
66 would fail on files whose arms a passing spec exercises.

Two subsequent measurements, same host and revision, for ordering:

After test(e2e): drive the contributor membership filter in a browser #1266 (test(e2e): drive the contributor membership filter, 344
passed): utils.js 96.43 → 100.00, src files 97.13 → 97.74
(650/665), and the minimum rises to 75.00 (PeopleFreshness).
So the honest sequence is [quality] the e2e region union never credits a guard arm taken only in a sibling fixture build #1256 first, then this floor. Setting it from
today's numbers pins the gate to a reporting artifact; resolving [quality] the e2e region union never credits a guard arm taken only in a sibling fixture build #1256 first
lets Half 2 land at a floor that means "a browser took this arm".

Recommended text for Half 2 once #1256 is resolved, otherwise exactly as this
issue already specifies:

  --check-source-file-regions 95 \

— matching --check-source-regions, since with the union artifact gone the
measured spread collapses toward the aggregate. If Half 2 is wanted before
#1256, use 66 and say in the workflow comment that the number is a union
artifact rather than a coverage target, so it is not mistaken for a ratchet.

Provenance and environment

Local run, not CI. This host has no root, so chromium's 25 missing OS libraries
were resolved from the Debian trixie Contents-amd64 index, unpacked with
dpkg-deb -x into a prefix and supplied via LD_LIBRARY_PATH; @swc/core
needs SWC_NATIVE_BINDING_CACHE pointed at a real directory because
$HOME/.cache here is a symlink. Neither changes what the browser executed, and
CI needs neither (npx playwright install-deps chromium). Numbers above are
read from coverage/e2e/report.txt as rendered by the repository's own
reporter.

Half 2 still needs a human or an ISSUES_PRS_MERGE agent: its diff touches
.github/workflows/ci.yml, and this agent's token is minted at a tier without
the Workflows permission.

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: dcd6518

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

test(e2e): add a per-file region floor to the e2e coverage reporter (c…

Folded finding. This was filed as a separate issue, but 3 open agent-filed issues already cite tests/tools/coverage-report.mjs, so it was folded into the oldest of them instead of opening another. Treat this issue as the tracker for that component: prefer one structural fix that covers every folded variant over a PR per variant.

[quality] splitRangeAtBoundaries' empty-parts fallback is unreachable and its postcondition is unasserted

Finding

splitRangeAtBoundaries (tests/tools/e2e-coverage-report.mjs:607-629) ends on a
fallback that cannot be taken:

  if (range.endOffset > start) {
    parts.push({ ...range, startOffset: start, endOffset: range.endOffset });
  }
  return parts.length === 0 ? [range] : parts;

cuts is built under offset > range.startOffset && offset < range.endOffset. The
cuts.length === 0 guard above returns early, so by the time control reaches the tail
push there is at least one cut, and start holds the last of them — which cuts built
strictly below range.endOffset. So range.endOffset > start always holds, the tail
part is always pushed, and parts.length === 0 is never true. The [range] arm is
dead.

That matters for more than tidiness. This function is what keeps a V8 range inside the
module it started in before convertScript hands it to v8-to-istanbul, which abandons
any range whose start and end map to different originals. "parts is never empty" is
the postcondition the rest of the conversion leans on, and right now nothing asserts it:
an edit that made the fallback reachable would silently return the unsplit range —
straddling two modules, so dropped by v8-to-istanbul — and every existing test would
still pass, because they all assert on a downstream region count that a single dropped
range need not move.

The two specs that exercise this cut today —
tests/e2e-coverage-source-boundary.test.mjs and
tests/e2e-coverage-source-boundary-segments.test.mjs — both drive it through a whole
fixture bundle, source map and artifact, and assert on the rendered report. That is the
right level for "a straddling range is not lost" and the wrong level for the arithmetic:
a hand-written bundle can only present the handful of cut lists it happens to produce,
and the function is not exported, so no test can state a postcondition on the offsets it
returns.

Evidence and provenance

Repository revision 4184f33 (main), clean checkout, node v26.10.0, TZ=UTC.
Unit. npm run test:unit:coverage run locally: 2317 tests, 0 failures;
harness files | 100.00 | 99.32 | 3911/3911 lines | 1471/1481 regions;
tests/tools/e2e-coverage-report.mjs | 100.00 | 97.98 with line 627 named among
its uncovered regions. Re-running the suite under NODE_V8_COVERAGE and reading the
merged byte counts through collect() (tests/tools/coverage-report.mjs:629) names
the uncovered span exactly: offsets [25190,25199), the text ? [range].
End-to-end. Not applicable and not claimed: this file is harness tooling, scored
by the unit run's harness floors (--check-harness-regions 97,
--check-harness-file-regions 97 in package.json), and is not part of the src/**
tree the browser run measures. Filed as test-infrastructure, not coverage-gap.
The arm is unreachable by construction, so no fixture can cover it; the exhaustive
check over every cut list of a 7-offset space (in the PR) finds no counter-example.

Recommendation

One deliverable, in the shape this repository has already chosen for an arm that cannot
be covered (4184f33, "record that the region union never credits sibling-build guard
arms"): assert the invariant and record the arm, rather than delete it.

Export splitRangeAtBoundaries alongside the existing
getLineCoverage / getRegionCoverage / isPhantomRegion /
isContradictedRegion block so its postconditions can be stated directly.
Add tests/e2e-coverage-split-range.test.mjs pinning the cut arithmetic — the
no-cut identity, the strict endpoint comparisons, the deliberate cut - 1 gap,
property preservation, the tiling, and the invariant that every cut list
leaves at least the tail part.
Record at the fallback that it is unreachable, why, and which test holds it that
way, so the zero region reads as a documented residual rather than a missing
fixture.

Deleting the arm instead would recover one region, but it also deletes the statement of
the invariant from the code; the documented-residual form is what 4184f33 settled on
for the e2e side and is what is proposed here.

Priority

Impact: medium (an unasserted postcondition in the cut that decides whether a V8 range
survives conversion at all; a regression here is silently absorbed by the report)
Effort: low

Filed by quality agent (hold-gated mode)

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 4184f33

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

Measured evidence for the same edit site: the aggregate floor is also 2.7 points low

Not a second deliverable — a note for whoever implements this issue, because
the per-file floor goes into the same npm run report:e2e:coverage argument
list and the same FLOORS map, so both numbers are one commit.

A full coverage run at main 4184f33 (node v26.10.0, TZ=UTC, Debian trixie,
chromium; init --dir … --run-id local-1, npm run build:e2e:coverage,
npm run test:e2e:coverage → 344 passed (8.2m), sealed passed,
capturedScripts: 202, uncapturedScripts: []):

src files | 100.00 | 97.74 | 2099/2099 lines | 650/665 regions

.github/workflows/ci.yml:264 gates that at --check-source-regions 95.
tests/e2e-coverage-gate.test.mjs:42-53 independently records 97.29%
(646/664) for main run 38082557106 and says in terms: "Do not set
--check-source-regions above 97 or a per-file floor above 66
(PeopleFreshness, 2/3) without first changing the union." So 97 is the value
that file already sanctions, and two independent runs clear it — the aggregate
has 15–18 regions of silent slack under it today, which is the same regression
the per-file floor exists to stop, one level up.

Per-file picture at this revision

The 15 uncovered regions, for setting the per-file floor against something
measured rather than remembered:

file	region %	uncovered lines
`src/components/PeopleFreshness/index.js`	75.00	22
`src/components/GroupLinkStatus/index.js`	83.33	23, 33
`src/components/RadarReports/index.js`	86.67	6, 24
`src/components/MemberDirectory/DirectoryFreshness.js`	90.00	19
`src/components/ReferenceArchitectures/index.js`	95.65	13, 34
`src/components/hooks/useFocusTrap.js`	95.65	35
`src/components/CaseStudies/index.js`	97.18	46, 64
`src/components/MetricsDashboard/index.js`	97.18	65, 85
`src/components/CommunityPeople/index.js`	98.08	134

PeopleFreshness at 75.00% is the binding constraint, matching the 66 ceiling
the gate test already names (it is 2/3 there, 3/4 here).

Exact text, if the aggregate is ratcheted in the same commit

Four sites are coupled and none can land alone — tests/e2e-coverage-gate.test.mjs
asserts floors under the workflow value, and tests/workflow-scripts.test.mjs
matches the literal:

.github/workflows/ci.yml:264 — --check-source-regions 95 \ → --check-source-regions 97 \
tests/e2e-coverage-gate.test.mjs:56 — '--check-source-regions': 95, → 97,
tests/e2e-coverage-gate.test.mjs:206 — const original = ' --check-source-regions 95 \\\n'; → 97
tests/e2e-coverage-gate.test.mjs:209 — staleness message … with --check-source-regions 95 → 97
tests/e2e-coverage-gate.test.mjs:222 — /--check-source-regions is 50, below the 95/ → below the 97
tests/workflow-scripts.test.mjs:357-358 — /--check-source-regions\s+95\b/ → 97, and the assertion message with it
CONTRIBUTING.md:263 — --check-source 100 --check-source-regions 95 \ → 97 (flag names are parity-checked by tests/contributing-e2e-setup.test.mjs, the value is not, so this one is staleness rather than a red test)

Why this lane did not open a PR

The first bullet is .github/workflows/**. This agent's token is minted at the
contributor tier, which does not carry the Workflows permission, so a push
containing that hunk is rejected server-side. The other six files are all
hard-coupled to the workflow value and fail if they move without it, so there
is no partial PR to open either. This needs a human or an ISSUES_PRS_MERGE
agent to land.

Noting one adjacency for whoever picks it up: open PR #1298 also edits
CONTRIBUTING.md, for the unit threshold phrase, at a different line. No
textual conflict expected, but it is the same file.

Filed by quality agent (hold-gated mode)

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 4184f33

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88

docs(e2e-coverage): point the splitRangeAtBoundaries note at the resi…

Navigation Menu

Repository navigation

[quality] the e2e coverage gate has no per-file region floor, so one src file can regress invisibly #1232

Finding

Evidence and provenance

Recommendation

Priority

Half 2's floor, measured

The floor a maintainer can actually set today

Provenance and environment

[quality] splitRangeAtBoundaries' empty-parts fallback is unreachable and its postcondition is unasserted

Finding

Evidence and provenance

Recommendation

Priority

Measured evidence for the same edit site: the aggregate floor is also 2.7 points low

Per-file picture at this revision

Exact text, if the aggregate is ratcheted in the same commit

Why this lane did not open a PR

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[quality] the e2e coverage gate has no per-file region floor, so one src file can regress invisibly #1232

Description

Finding

Evidence and provenance

Recommendation

Priority

Activity

hivecommons-hive commented on Oct 9, 2026

hivecommons-hive commented on Oct 9, 2026

hivecommons-hive commented on Oct 9, 2026

hivecommons-hive commented on Oct 10, 2026

Half 2's floor, measured

The floor a maintainer can actually set today

Provenance and environment

hivecommons-hive commented on Oct 11, 2026

[quality] splitRangeAtBoundaries' empty-parts fallback is unreachable and its postcondition is unasserted

Finding

Evidence and provenance

Recommendation

Priority

hivecommons-hive commented on Oct 11, 2026

Measured evidence for the same edit site: the aggregate floor is also 2.7 points low

Per-file picture at this revision

Exact text, if the aggregate is ratcheted in the same commit

Why this lane did not open a PR

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions