Repository navigation
Description
Gates 52, 68 and 69 each carry a ratchet. None of the three ratchets runs in CI. They have not run in any repository since full scope became the default.
The mechanism
All three hand their base ref to their checker only inside the same condition:
if [ "${SCOPE_TO_DIFF}" = "1" ]; then
HYDRA_GATE_BASE_REF="${BASE_REF}" python3 .../check_x.py
else
python3 .../check_x.py # no base
fiThe shared quality workflow never puts the runner in diff scope. --scope-to-diff appears in no workflow file, and full scope has been the default since ADR-020 was superseded. The else branch is the only branch CI takes.
Each checker then skips its ratchet, because a ratchet with no base has nothing to compare against.
Measured
dossiq PR #1985 added a type:"custom" page and took the app from 11 to 12. From that run's own log:
[hydra-gates] Delta base: origin/development ($HYDRA_GATE_BASE_REF) = fa09f5837
[hydra-gates] SCOPE-MODE: full
[gate-69] page-type-discipline: PASS
The delta base was resolved. The gate's own wiring threw it away and printed PASS.
Reproduced locally on dossiq at development, by planting a thirteenth type:"custom" page and running gate-69's checker both ways:
| how it was run | result |
|---|---|
no HYDRA_GATE_BASE_REF, the mode CI uses |
0 findings, empty log |
HYDRA_GATE_BASE_REF=HEAD |
custom-page ratchet: 12 -> 13 (+1) |
dossiq's custom-page count rose 11 to 12 twice since gate-69 landed on 2026-08-21, in #1677 and #1985. Neither run reported anything.
Why it is worse than it looks
Gate-52 already prints a line saying its ratchet half was not computed, added in #374. Gates 68 and 69 say nothing at all. A gate that is registered, prints PASS, and never ran reads exactly like a gate that passed.
Gate-69's ratchet is the one ADR-100 is enforced by, so ADR-100 has no mechanical enforcement today.
The fix, and why it is not a one-liner
Five gates already take their base off the delta channel regardless of file scope: 16, 29, 47, 48 and 61. The runner's own header describes that as the design. Moving 52, 68 and 69 into that family is a few lines each.
What makes it a decision rather than a patch: all three are blocking. The day they connect they start firing fleet-wide on growth that has been landing unmeasured for weeks. That wants measuring first, per repo, and probably a warning period.
Gate-114 (#723) is wired the correct way and carries the arm that would have caught this: with no base, the run names the half that did not run rather than printing a bare PASS.
What to do
- Measure what each of the three would report today, across the 21 core apps.
- Decide per gate whether it reconnects blocking or advisory.
- Reconnect, and give each one gate-114's arm 6, so a ratchet that cannot compute says so.
Activity