|
| 1 | +# 150 — outcome: what shipped, what stayed open, what is still unverified |
| 2 | + |
| 3 | +Closing record for the 260819 unclaimed-bug unit. Five PRs on `dev`. |
| 4 | + |
| 5 | +| PR | dev | issue | disposition | |
| 6 | +|---|---|---|---| |
| 7 | +| #2116 | `bdaf4463a` | #2107 | **closed** — proxy env baked into service definitions | |
| 8 | +| #2117 | `c701cc7e6` | #1933 | open — decoding fixed, diagnosis unproven | |
| 9 | +| #2118 | `239cae9e5` | #1527 | open — 1 of 5 residuals fixed | |
| 10 | +| #2121 | `fbc6f26a2` | #2108 | open — diagnostic shipped, mechanism not | |
| 11 | +| #2126 | `8e7b63387` | — | audit remediation | |
| 12 | + |
| 13 | +Only #2107 is closed, and that asymmetry is the point: three of the four fixes |
| 14 | +are slices, and saying so on the issue is cheaper than a reopen. |
| 15 | + |
| 16 | +## The two defects our own work introduced |
| 17 | + |
| 18 | +Worth leading with these rather than the wins. |
| 19 | + |
| 20 | +**A cross-file test leak that read as CI flake.** The #2107 tests assigned proxy |
| 21 | +variables onto the real `process.env` and restored them in a `finally`. `bun |
| 22 | +test a b` runs every file in ONE process — `--isolate` does not change that — so |
| 23 | +the values outlived the file, and the Lab sandbox rejects any live proxy variable |
| 24 | +as `harness_failure`. 73 macOS failures, zero when the Lab suites ran alone. |
| 25 | + |
| 26 | +**A credential written world-readable.** #2107 made service definitions |
| 27 | +credential-bearing (a proxy URL routinely carries `user:password`) while they |
| 28 | +were still written with a bare `writeFileSync` — 0644 under umask 022, measured. |
| 29 | +In a file where the API token and install state already use 0600, and in a repo |
| 30 | +whose `collectProxyEnv` deliberately reports proxy presence as a boolean so the |
| 31 | +URL never escapes. |
| 32 | + |
| 33 | +The second one merged. It was caught only because the final audit asked a |
| 34 | +question the author never asked — *what mode is that file?* — and then measured |
| 35 | +instead of reasoning. |
| 36 | + |
| 37 | +## What actually caught things |
| 38 | + |
| 39 | +- **Ablation, three times.** The #1527 fix looked right at `failAndClear` and all |
| 40 | + ten tests still passed — that helper is not on the failure path. The first |
| 41 | + proxy-leak theory (Bun spreading `null`) produced 50 fail / 50 fail with and |
| 42 | + without the "fix". A number that does not move is not a fix. |
| 43 | +- **Running the suite the way CI runs it.** Per-suite local green missed the leak |
| 44 | + entirely, because the defect only exists across suites in one process. |
| 45 | +- **Removing our commits.** The same two files on `origin/dev` went 144 pass / 0 |
| 46 | + fail. That is what converts "CI is flaky" into "we broke it". |
| 47 | +- **An auditor that re-derives rather than re-runs.** Round 1 returned fail on |
| 48 | + #2121's design; the final round found the P1 and mutation-tested our |
| 49 | + assertions in a scratch tree to prove they were not vacuous. |
| 50 | + |
| 51 | +## Still unverified, stated rather than filed as done |
| 52 | + |
| 53 | +Everything below is real and none of it is closed by this unit. |
| 54 | + |
| 55 | +- **Windows.** The Windows ACL half of the 0600 fix and the `utf16le` scheduler |
| 56 | + assets are asserted for mode on POSIX only. `owner-unavailable` — the branch |
| 57 | + #2108 most needs named — is a Windows icacls path asserted nowhere in the |
| 58 | + suite. |
| 59 | +- **#1933's diagnosis.** The encoding mechanism is proven in code; that it is |
| 60 | + *this reporter's* problem is inference from a GitHub display name. If their |
| 61 | + path is ASCII, the diagnosis is wrong and the issue should stay open. |
| 62 | +- **#1527's other four residuals.** kimi-k3 collapse and the 429 asymmetry depend |
| 63 | + on #2054. The 429 half may be unprovable while Connect withholds |
| 64 | + `cache_read_tokens`. |
| 65 | +- **#2108 phase 2.** Waiting on a field report that names a reason. Two candidate |
| 66 | + triggers, and phase 2 targets one of them. |
| 67 | +- **wp1587**, the deferred tool catalog, was never implemented. Both its blockers |
| 68 | + cleared, so it is ready, not done. |
| 69 | + |
| 70 | +## Final state |
| 71 | + |
| 72 | +`dev` full suite on `ssh lidge`: **13,505 pass / 15 skip / 0 fail** across 855 |
| 73 | +files. Cross-platform CI green on the merged head. |
| 74 | + |
0 commit comments