[noidee-dev]

Follow-up to #8, which moved the gate to macos-latest and left three things owed: a step budget resting on a figure stitched from two runs, two citations naming heads the merge rebase rewrote, and a backlog row whose next step that PR had quietly fulfilled.

The budget: 16 -> 20 minutes

Withdrawn, quoted verbatim: "16 minutes (960 s) is 2.04x of 392 s + 78 s = 470 s -- this image's measured compile plus the OTHER image's measured run, because this image's run figure does not exist yet: the step was killed before the suite finished."

That paragraph promised its own replacement by the first green run on this image. Keeping the promise exposed the shape of the error rather than just the missing addend: a step budget bounds a step's wall clock, and adding one run's compile to another run's suite duration bounds a step that neither run is. That wall clock is a field pair of the Actions API, present on every run that reached the step:

run	image	Unit-Tests step	conclusion
38225383495	macos-15	329 s	success
40933631584	macos-latest	493 s	failure — the 480 s budget's own kill
41045404989	macos-latest	581 s	success
41047833837	macos-latest	517 s	success
41048445836	macos-latest	640 s	success

640 s is the maximum of a step that finished, so 1200 s is 1.88x — the first of this comment's derivations to land under 2, after 2.09x and 2.04x. Neither of those states 2x as a requirement, so the comment now says in as many words that "the 2x floor" is a reading of them and not a quote, and names 22 minutes (1320 s, 2.06x) as the smallest whole-minute budget holding that reading. 20 is the maintainer's decision of 2026-10-31, taken while the best figure on offer was the 558.79 s the withdrawn stitching produced, against which 1200 s read as 2.15x. The series was measured after the decision, not before it.

The spread matters more than the maximum, and one pair isolates it: 517 s against 640 s is one commit apart, and that commit adds 23 lines of CLAUDE.md, which no target compiles. The compile half alone moved 345.71s -> 492.86s across it.

The job budget is unaffected: 753 s and 937 s for the two green runs against its 2100 s, and a step running the full 1200 s would put the worse of those at 1497 s, 0.71x.

Two heads the rebase rewrote

Both citations now name the develop commit beside the head the run actually had. Replacing the hash alone would have claimed a run ran on a commit it never saw. git show <h> | git patch-id --stable gives one id per pair: 9c8d3673/a774ac4d in ci.yml, c8041d19/bcd0bc61 in ConnectMainActorLivenessTests.

docs/BACKLOG.md

Two rows corrected in place with their wording quoted, two opened.

The row "Every layout figure the guards hold was measured on Darwin 27" asked for "the first push or pull request that triggers a run, and then reading the checks' verdict". #8 was that run, and it answers half the row: both green runs passed Suite "Bounded view layout guard" and Suite "Resizable container floors" on macOS 26.6.2, so the local Darwin 27 figures survive a Darwin 26 runner across all 33 bounded views and the resizable floors. By how much is still unmeasured — both reporters are gated, so the runs produced a verdict and no figures, and a pass is not a margin. The macos-15 premise is withdrawn with the quote, as is the now-false title half of the row it was split from.

Opened: the three-way toolchain split (gate on Swift 6.3.3, the maintainer's Mac on 6.4, the retired image on 6.1.2), whose first instance was the allLines warning that reddened run 41045404989 at 26 occurrences on a tree every macos-15 run had built clean. And claude/nice-leakey-5da0b9, surfaced by a local branch sweep in which 26 of 32 local branches were content-identical to develop or main and were deleted: it carries ed0b70af, a finished relaunch fix dated 2026-09-05 that is on neither develop nor main.

Verification

Every command written into the comment and into the rows was extracted from the committed text and run in that form; all six reproduce their stated figures. Two false claims were caught in the new comment before committing — "two commits apart" (it is one) and "one test-function annotation" (it is 23 lines of Markdown) — and the correction strengthens the variance argument rather than weakening it.

No local suite run: the only Swift change is four doc-comment lines, parse-checked with swiftc -parse against a positive control that reported 2 errors on a broken file. This PR is what tests the new budget — ci.yml triggers only on a push to main/develop and on pull_request, so the branch alone starts nothing.

---

20fbe7f2 — the first run on the new budget went red, and not on the budget

Run 41244948849 took 597 s for Unit-Tests, comfortably inside the new 1200 s, so nothing was killed. It went red on one test: ConnectionDiagnosticsTests.theSSHDialNeverPutsTheSecretInTheReport recorded "a closed port did not fail the SSH dial: .timedOut". The dial never failed — its step was settled by its own deadline before the probe was given a thread. The case handed the runner stepTimeout: .seconds(10), and in that run 8184 of the 8307 cases carrying a duration line reported a wall clock at or above 10 s, against a run total of 116.701 s.

That is docs/BACKLOG.md's "Wall-clock ceilings still in the tree", fifth shape, reaching the test through a production parameter the test itself chooses. It had already been bought one file at a time, twice: 61850a9c (2026-09-19) moved the jump suite's twin off .seconds(10) for exactly this reason — one file, 10 insertions, 2 deletions — and ResolveNamesTests had already named its own 600 s roomy. The five literals in ConnectionDiagnosticsTests were never swept and fell 42 days later.

• A named roomy of 600 s carrying the measurement and the command that produces it, with all five call sites through it. Raising it costs those five nothing, which is the argument that makes 600 s safe here: every target either answers at once (the rig) or refuses at once (a closed loopback port), and the skipped case never dials.
• Both diagnostics suites withdraw the clause "it is deliberately far larger than any step budget below", quoted in place. False of this file by construction now; false of the jump file on the day it was written — git blame dates both of that file's 600 s sites to 61850a9c and the sentence's heading to ae19a53a, nine days later.
• Deliberately not swept, with the command that lists them: six occurrences in ICMPEchoTests and NetworkTraceTests. Five send real echo or trace probes, where a roomy budget's cost is unmeasured; the sixth is NetworkTraceTests.traceStep, whose own comment calls 5 s generous and records the 5.004 s behind it. Two more route through per-file factories shared with cases where a tight budget may be load-bearing. Each needs classifying before any value moves — which is how 2026-09-19 produced this red.

Red first, in two parts

The observed red is the CI run. The mechanism reproduces at a keyboard with the budget planted at .nanoseconds(1), where the deadline cannot lose: the SSH case recorded the identical sentence, and the S3 case its own twin. Two issues in that run and three in the first one, same plant — which case varied is not established, because the earlier run's per-case lines were lost to a tail in the command that captured it. Green at 600 s: 154 tests in 6 suites, 0 issues.

One pre-existing red found by the full local run, not fixed here

ViewTestabilitySpike's "Varying only isRegex leaves the pixels unchanged" fails on this Mac (macOS 27.0.1, Swift 6.4.2) and passed on the gate in run 41244948849 itself, after 1.265 s. The case's doc comment states as measured that a Toggle's checked state "stays out of the bitmap"; on macOS 27 it does not. Recorded in the toolchain-split row, which withdraws its own "nothing has been run to find what Swift 6.4 diagnoses that 6.3.3 does not" as far as "nothing has been run" goes — what was found is a rendering divergence, not a diagnostic one, so the row stays open. Untouched here by construction: the spike is in macSCPAppKitTests and this commit touches only macSCPCoreTests and docs/.

Full local suite on the committed tree: 5091 tests in 408 suites passed, 2884 in 242 suites with that one pre-existing issue.

🤖 Generated with Claude Code