Gladex Agent Logs

  orphan by construction) and comes back **orphan alive + waited child gone**;
  the **(86)** shape is run end to end too (**B14–B17**): the reader's verdict
  on that very file is *not* waited, the fixture is done in **1 s rather than
  61**, the job it named **is** reaped, and the unnamed sibling is **still
  alive** — a process that falsifies the pre-(86) verdict rather than a
  comment that asserts it; and everything the suite started is gone after the
  reap. Since `[0.4.152]` the (101) shape is run end to end too (**B18–B20**):
  the reader's verdict on the `wait;` file is *waited*, the fixture exits **0**
  (its wait really completed instead of orphaning the child), and the child it
  waited for is **gone**. Since `[0.4.156]` the (104) shapes are run end to end
  as well (**B21–B26**), in both directions at once: the reader's verdict on
  the `&&` file is *waited* and the process agrees (fixture exits **0** in ~1 s,
  child gone), while its verdict on the pipeline file is *not* waited and the
  process agrees too (fixture exits **0 immediately**, `sleep 61` still alive
  and asserted so) — the measured half of the subshell rule, without which the
  widened start class would have bought a false positive on that very file.
- **C (3)** the rule rather than the instance: no tool creates a foreign
  session without owning the cleanup for it, exactly one tool needs that guard
  today, and no fixture process outlives the suite.
- **K (68)** the seventeen mutation controls below, **executed** — queue item
  **(100)**, `[0.4.151]`, two more added by **(101)**, `[0.4.152]`, and two
  more by **(104)**, `[0.4.156]`.
  Every figure this section has recorded since
  `[0.4.126]` used to be prose: `grep -rn "test_detached_children" tests/
  tools/` found exactly two hits (the suite and this section), the file had no
  in-file control section, and the patches were last re-run *by hand* in
  sandboxes on 2026-09-30 — evidence, not automation, so a reader changing a
  plant would rot the table in silence. Section K runs all seventeen on every
  invocation in four steps, **`Kn a`–`Kn d`** each: plant ONE mutation in a
  fresh copy of `tests/` + `tools/` under `$SB/ctl/` (the live tree is never
  written) and assert it **applied** — a plant that did not take comes back
  0 red, the exact mistake `[0.4.126]` made twice; assert the child ran the
  **whole** 70-assertion suite (`passed + failed == 70`, so a child that died
  early cannot be read as "few reds"); assert its `FAIL` labels are **exactly**
  the documented set **in order** (not "at least these", so a mutation that
  reddens something nobody predicted fails too); assert the exit code is `1`
  when something is red and `0` for the untouched copy. The child runs with
  `DETCHILD_CONTROLS=skip`, so a control cannot recurse into itself — that
  skip is what keeps every child at 70. Two of the three plant helpers were
  broken by their own section when it was written and the plants caught them:
  `awk -v sub=` dies in gawk (`sub` is a builtin) and the section quotes every
  pattern it plants, so the plants' `exactly one match` guard counted the
  quotes — both are why `Kn a` exists rather than trusting the mutation. The
  third helper defect was caught the same way in `[0.4.153]`: `awk -v pat=`
  **interprets backslash escapes**, so K14's needles (`wait([ \t;|]|$)`) were
  passed as a real TAB, matched nothing, and the plant refused — the helpers
  now carry `pat`/`rep` in `ENVIRON`, where no escape is read.

**Mutation controls** (since `[0.4.151]` these are **asserted by section K**,
one `Kn a`–`Kn d` per row, on every run; the figures were first measured by
hand in sandbox copies of `tests/` + `tools/`, live tree never written, every
plant asserted to have applied before the child suite is started): a second
tool gaining `setsid` → **3 red**
(`A2 C1 C2`), a second tool gaining a background `&` → **1 red** (`A4`), the
real tool's `wait` commented out → **1 red** (`A5`), the guard's tokens renamed
to tokens that do **not** contain the searched one → **2 red** (`A3 C1`), and
an untouched copy → **0 red / 42**. Two more, one per direction of the
`[0.4.127]` (84) fix: the **pre-fix last-vs-last reader** installed back →
**1 red (`A15`)**, and an **over-strict pairing** where a bare `wait` covers
only one child → **1 red (`A16`)**. **All seven were re-run and re-measured on
2026-09-30 at 42 assertions** — the six historical figures hold unchanged, and
the untouched copy is 42/0 (section K's K5 measures it at **0 red / 70**, the
child's own count since `[0.4.156]`; it read 56 from `[0.4.152]`, 50 from
`[0.4.145]` and 42 before that).
- **Five rows re-measured and two controls added since `[0.4.156]` (queue item
  (104)) — all seventeen run at 138 assertions, measured 2026-10-01, child
  70**: the **pre-(84) last-vs-last reader** now reads **9 red
  (`A15 A27 A28 A29 A38 A39 A40 B14 B24`)**, 61/9 — it asks only whether a
  WAIT record follows the last `&`, and a pipeline's wait prints one, so it has
  no subshell rule to lose; the **over-strict pairing** now reads **3 red
  (`A16 A32 A37`)**, 67/3 — A37 is the third instance of its own subject, a
  bare wait over three children, this time closed by `&&`; the **pre-(86)
  counting reader** now reads **9 red
  (`A26 A27 A32 A37 A38 A39 A40 B14 B24`)**, 61/9 — neither the operator cut
  nor the subshell rule was ever in it; the **operator cut removed** now reads
  **3 red (`A29 A37 A40`)**, 67/3 — the cut is also what sets `op`, i.e. what
  flags a left-hand subshell wait, so A40 is the cut's own casualty and not a
  stray; and **K14** now reads **6 red (`A30 A31 A32 A35 A37 B18`)**, 64/6 —
  A35/A37 join because their waits are closed by `&`, the terminator K14
  narrows away. **K15 stays at 1 red (`A32`)**, 69/1: A37 does **not** join it
  (measured, not assumed — the operator cut fires on the leading `&` whether
  the tail hands it back or not), which is why the `&` half of the hand-back
  class is queued as **(107)** instead of being documented as covered. The two
  new controls: **K16** puts both live readers back on the pre-(104) class →
  **6 red (`A33 A34 A35 A36 A37 B21`)**, 64/6, and **K17** deletes the subshell
> build · mimo-v2.6-flash-free
$ cd /data/repo && sed -n '3730,3840p' tools/REGISTRY.md
**6 red (`A33 A34 A35 A36 A37 B21`)**, 64/6, and **K17** deletes the subshell
  rule → **4 red (`A38 A39 A40 B24`)**, 66/4. The ten rows neither changed nor
  added here (**K1–K5, K8, K9, K11, K12, K15**) are unchanged at their
  measured sets, and the untouched copy is **0 red / 70**.
- **Two rows re-measured and two controls added since `[0.4.152]` (queue item
  (101)) — all fifteen run at 116 assertions, measured 2026-10-01**: the
  **over-strict pairing** now reads **2 red (`A16 A32`)**, 54/2 — a bare
  `wait` closed by a separator over three children is the same subject and A32
  is its second instance, which the control predicted before it was run; the
  **pre-(86) counting reader** now reads **4 red (`A26 A27 A32 B14`)**, 52/4 —
  that reader never had (86)'s operator cut either, so `wait; echo done` reads
  `; echo done` as three arguments, `nargs ≠ 0`, one decrement over three
  children, two left pending; the cut that would stop the list at the separator
  is (86)'s own subject (K13), which is why the fourth red is documented rather
  than tidied away. The two new controls: **K14** puts both live readers back
  on the space/EOL terminator class → **4 red (`A30 A31 A32 B18`)**, 52/4, and
  **K15** lets the tail swallow the separator again → **1 red (`A32`)**, 55/1.
  The thirteen older rows are unchanged at their measured sets.
- **Four new controls from `[0.4.145]`, and all thirteen re-run at 50
  assertions the same day, untouched copy 50/0**: the **pre-(86) counting
  reader** installed back → **3 red (`A26 A27 B14`)**, 47/3 — arity fails a
  correct tool, identity passes a wrong one, and the live fixture's verdict
  goes with it; **identity resolution removed** (an argument always covers the
  oldest uncovered job) → **2 red (`A27 B14`)**, 48/2; **arity removed** (one
  cover per `wait`, whatever its list names) → **1 red (`A26`)**, 49/1; the
  **operator cut removed** → **1 red (`A29`)**, 49/1, because `return 1` is
  then read as two more job names. The nine older rows re-measured at 50: the
  seven that do not involve the reader are unchanged (**3 / 1 / 1 / 2 / 1 /
  2 / 2 red**), and the **pre-(84) last-vs-last reader** now reads **5 red
  (`A15 A27 A28 A29 B14`)**, 45/5 — one row *changed*, because every (86)
  plant and B14 put a `wait` after the file's last `&`, which is the only
  question that reader asks. The 42-assertion figure above is left as
  recorded; 5 is the measured number now.
- **Two new controls from `[0.4.144]`, both measured the same day**: the
  `cmd-str` rule removed → **2 red (`A18 A23`)**, 40/2 — the code-as-string
  call site goes unreported and the live non-vacuity goes with it; and the
  anchored `source` rule replaced by the naive prefix match → **2 red
  (`A19 A25`)**, 40/2, with the live tree reading **10 real + 8 prose** call
  sites under that rule — which is the number A25's text claims, measured from
  the mutation rather than guessed beforehand.
- **Correction to `[0.4.126]`'s table, re-measured in `[0.4.127]`**: that entry
  recorded the `setsid` row as **4 red (`A2 A3 C1 C2`)**, and it is **3**: A3's
  predicate inspects `tools/regression-run` only, which the "second tool gains
  `setsid`" plant does not touch, so A3 cannot fire — the same file's own
  comment already says so (*"a second detacher fails C1 rather than making A3's
  own subject ambiguous"*). The row above is the measured one; the historical
  entry is left as written.
- The first version of the plants failed twice and both failures were in the
  *plant*, not the suite: `: # wait …` still read as a wait (fixed by
  comment-stripping in the reader, now plant **A12**), and
  `_LIVE_PGIDS_PLANTED` still matched `grep -q _LIVE_PGIDS` (fixed by renaming
  to a token that does not contain the searched one). **Both mistakes were made
  again while re-running the controls in `[0.4.127]`** — `: wait "$pid"` read
  as a wait and `_LIVE_PGIDS_PLANTED_X` still matched — and came back as **0
  red**, i.e. a green that proved nothing; fixed by using `#`-comment form and
  non-containing tokens (`LIVE_GROUP_SET` / `install_guard_handler`), after
  which they read `A5` and `A3 C1` as recorded.

**Since `[0.4.156]`** — queue item **(104)** widened the class again, on both
sides of `wait`: a wait whose NEIGHBOURING character is `|` or `&` is a wait
too, so the census's WAIT rule and the pairing's match both read
`(^|[; \t|&])wait([ \t;|&]|$)` and the tail hands `&` back the way it hands
back `;` and `|`. Three plants cover what that buys — `true&&wait`
(**A33**/**A36**), `true|wait` (**A34**/**A38**) and `wait&&echo done`
(**A35**/**A37**) — and three cover what it must NOT buy: bash runs a
pipeline's `wait` in a subshell with an empty job table, so `true|wait`,
`true | wait` and `wait | true` read as **not** waited (**A38**–**A40**,
**B24**), a rule the pairing carries as `head ~ /\|[ \t]*$/ || op == "|"` with
**A39** the false positive this step *closes* (the spaced spelling already
matched the pre-(104) class and read as waited). Two controls take the two
premises away — **K16** the class, **K17** the subshell rule — five rows'
documented sets were re-measured rather than inherited, and the `&` half of the
hand-back class turned out to have **no shape that reddens it** (K15 reddens
on `A32` alone), so that is written into the reader's own comment and queued as
**(107)** instead of being claimed as covered. Assertions **116 → 138**
(**A 33 → 41, B 20 → 26, K 60 → 68**, `C 3` unchanged), the child suite
**56 → 70**, controls **15 → 17**, runtime ≈ 3 m 6 s → ≈ 4 m 6 s, suite count
unchanged at **75**.

**Since `[0.4.152]`** — queue item **(101)** widened the `wait` terminator
class both A-section readers match against to `(^|[; \t])wait([ \t;|]|$)`, so
a wait closed by `;` or `|` is a wait — the census's WAIT record (`A30`) and
the pairing's verdict (`A31`, `B18`) both see it — and the `wait` line's
**tail hands the separator back** instead of consuming it, so `wait; echo done`
stays a bare wait over all three of its children (`A32`, `B19`/`B20` measure
the same file as a process). Two controls take that premise away (**K14**,
**K15**), and `[0.4.153]` corrected the control machinery the first full run
reddened on: the plant helpers carry their needles in `ENVIRON` rather than
`awk -v` (which reads `\t` as a TAB, so K14 could not plant), only `;` and `|`
are handed back to the tail (a whitespace terminator left a leading empty
token for `split`, which made K12's capped loop count nothing and read every
argument-bearing wait as bare), and K10's fourth red was written down instead
of designed away. Assertions **102 → 116** (**A 30 → 33, B 17 → 20, K 52 →
60**, `C 3` unchanged), the child suite **50 → 56**, controls **13 → 15**,
runtime ≈ 2 m 30 s → ≈ 3 m 6 s, suite count unchanged at **75**.

**Since `[0.4.151]`** — queue item **(100)** gave the thirteen mutation
controls above an in-file home: section **K** runs them, so the table is read
back by the same file that writes it. Assertions **50 → 102** (**A 30, B 17,
C 3, K 52**), runtime ≈ 10 s → ≈ 2 m 30 s, the child suite still **50** (K is
skipped there by `DETCHILD_CONTROLS=skip`), the suite count unchanged at
**73**, and no figure in this section's control table changed: every one of
the thirteen measured the reds it was written down as measuring.

**Since `[0.4.145]`** — queue item **(86)** replaced `bg_waited`'s counter
with **job identities**: each `&` is a numbered job, `name=$!` records which
job a name refers to, `wait` covers one job per argument resolved by that
identity (falling back to the oldest uncovered job when the name cannot be
resolved), the argument list is cut at the shell operator that ends it, and a
bare `wait` still covers everything started before it. Plants **A26–A29**
(arity, identity, arity's ceiling, the operator cut) plus the live pair
> build · mimo-v2.6-flash-free

Select another run log from the list above. Only files matching run-YYYYMMDD-HHMMSS-N.log are readable.

[Thu Oct  1 20:49:35 2026] 127.0.0.1:36036 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36036 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36044 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36044 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36052 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36052 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36056 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36056 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36060 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36060 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36076 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36076 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36092 Accepted
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36092 Closing
[Thu Oct  1 20:49:35 2026] 127.0.0.1:36094 Accepted
[Thu Oct  1 20:49:36 2026] 127.0.0.1:36094 Closing
[Thu Oct  1 20:49:36 2026] 127.0.0.1:36096 Accepted
[Thu Oct  1 20:49:36 2026] 127.0.0.1:36096 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56406 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56406 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56412 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56412 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56424 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56424 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56426 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56426 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56434 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56434 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56440 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56440 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56448 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56448 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56460 Accepted
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56460 Closing
[Thu Oct  1 20:50:20 2026] 127.0.0.1:56464 Accepted
[Thu Oct  1 20:50:21 2026] 127.0.0.1:56464 Closing
[Thu Oct  1 20:50:21 2026] 127.0.0.1:56466 Accepted
[Thu Oct  1 20:50:21 2026] 127.0.0.1:56466 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60342 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60342 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60356 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60356 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60364 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60364 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60378 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60378 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60392 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60392 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60406 Accepted
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60406 Closing
[Thu Oct  1 20:51:54 2026] 127.0.0.1:60418 Accepted
[Thu Oct  1 20:51:55 2026] 127.0.0.1:60418 Closing
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52126 Accepted
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52126 Closing
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52142 Accepted
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52142 Closing
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52146 Accepted
[Thu Oct  1 20:51:55 2026] 127.0.0.1:52146 Closing
[Thu Oct  1 20:52:51 2026] 127.0.0.1:37434 Accepted