Skip to content

The parse gate has a false-green mode, and this document caused it: the indexing line is not the verdict #8997

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
gunbai-bot wants to merge 2 commits into from
Closed

The parse gate has a false-green mode, and this document caused it: the indexing line is not the verdict #8997

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a5866a9
The parse gate has a false-green mode, and this document caused it: t…
Aug 23, 2026
1c2e851
Merge main into docs/parse-gate-false-green: the three width/DIMM rep…
Aug 23, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 56 additions & 4 deletions docs/probes/local_dag_parse_gate_recipe_2026-08-22.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,12 +29,60 @@ module index refused: 1 unparseable .dag source(s)
dag/gunbc/ci_layer_roots.dag:22395-22396: expected expression, found Slash
```

## The indexing line is NOT the verdict — wait for the exit code

**This is the way to get a false green, and a false green here is worse than no gate**, because you
will have "run the check". Added 2026-08-23 after `neat-heron-312` came within a minute of reporting
the recipe as broken, and after this document's own earlier claim about *when* the check fires turned
out to be wrong.

The run prints `indexed 3884 modules from 2 source roots` and **succeeds** at that step. The
annotation-grain and parse diagnostics do not land there: they arrive after reconcile, at the very
end — measured at **128 seconds** on one box. Someone who starts the command, sees a clean indexing
line at 95 seconds, and interrupts will read a defect-free tree that is not defect-free.

So the verdict is `EXIT=0` plus `0 diagnostics` on the final line. Nothing earlier counts.

```
... ; echo "EXIT=$?"
```

**The mechanism claim in the next section is corrected by this.** An earlier revision of this
document (and a message this session sent to two other lanes) said annotation grain is checked
*during indexing*. It is not — indexing reads the sources, and the refusal is raised later. What
survives unchanged is the part that makes a trivial entry sufficient: the diagnostic is raised for
modules the entry never imports and never compiles.

## The gate has been run in both directions

A check that has never gone red is a decoration. `neat-heron-312` ran both arms on one tree with one
binary (2026-08-23):

| arm | result |
|---|---|
| in-body `//` planted in `dag/product/workload_simulation.dag` | `EXIT=1`, `1 hard diagnostic`, located to that file |
| reverted, same command, same binary | `EXIT=0`, `compiled: 6 files emitted, 0 diagnostics`, 128s |

The planted defect was in a module the trivial entry does not import. **Six files compiled, 3884
indexed, and the diagnostic came from one of the 3878 that were never compiled at all** — which is
the load-bearing evidence for the section below, and stronger than the reasoning that motivated it.

## A stale binary fails loudly and looks like a catastrophic finding

If your binary predates the `--entry` flag it refuses the flag outright. Falling back to a
whole-corpus compile *without* an entry selects a parser that rejects `//` entirely and returns
**~25964 errors on a clean tree** (`warm-tern-755`, 2026-08-23, confirmed against a stashed HEAD as
baseline noise).

That is the absorbing fallback with the failure wearing the costume of an answer: the degraded mode
is loud, and its output reads as a discovery rather than as a broken instrument. **Five-figure error
counts mean check your binary, not check the corpus.**

## Why a trivial entry suffices

The entry selects what gets **compiled**; it does not bound what gets **indexed**. Building the module
index reads every `.dag` source under every declared root before any entry closure is resolved, and it
refuses on any unparseable one. So the parse phase is corpus-wide *by construction* and a two-line
entry pays only the index, not a real compile — which is what makes it a minute rather than the eight
The entry selects what gets **compiled**; it does not bound what gets **read**. Every `.dag` source
under every declared root is read and checked regardless of the entry closure, so the parse phase is
corpus-wide *by construction* and a two-line entry pays only that sweep, not a real compile — which is what makes it a minute rather than the eight
that a floor run costs.

That also fixes what the recipe does and does not tell you. It answers exactly one question — does
Expand All @@ -43,6 +91,10 @@ not regen, not the witness floor, and a green run here says nothing about any of

## Reading the coordinates

The local gate locates by **character offset**; the floor's message for the same class gives
`file:line:col`. So this tells you the file and makes you hunt within it — not a defect, but do not
assume your grep is broken when the number does not match a line.

`file:START-END` is a **byte** span, not a line range. `sed -n '1,NNNp'` on those numbers will point at
the wrong place. To see the offending text:

Expand Down
Loading