Two declared corrections: fabric_evidence_gating's deleted-script claim, and the FCI-1 runners' (c) class

[gunbai-bot]

Correct gunbc.rung_drop fabric_evidence_gating: the calibration script was restored, the lane was not

The row's 2026-09-04 amendment says that ruling deleted tools/fabric_ci_evidence_calibration.sh. Git history shows two separate facts:

• Cut three of seven CI jobs, and close the roster to growth without operator sign-off #10390 (2026-09-04) deleted it along with the fabric-evidence job, the calibration step and its bounds.
• FCI-1: bind exact Work to a durable converged-cell reservation (supersedes #9693, re-derived against current main) #10270 (FCI-1, 2026-09-05, "re-derived against current main") re-added it. It is live: tools/fabric_ci_fci1_live_instrument.sh invokes it against the exact binary before any srv3 observation.

So the row overstates, and the script is not dead. Deleting it would break the FCI-1 live instrument.

Declared correction. A CORRECTED 2026-09-28 paragraph now opens the row. It quotes the wrong clause instead of deleting it, which follows the existing correction pattern (e.g. floor_cut_heal). It then states:

• Changed: the script exists and runs, but only when an operator invokes the FCI-1 live instrument by hand, on no CI trigger.
• Unchanged: the job, the step and the FABRIC_EVIDENCE binding are still absent from every emitted workflow, and w_RED_the_deleted_lanes_do_not_return still guards that. The rung, the population and the restoration trigger stand.
• The script's own scaffold and dissolution header are a separate subject that this row does not govern.

docs/design-rung-drops.md is regenerated through tools.docs_projection_gate regen, not edited by hand. The diff is that one line.

Part of node adhoc-da321263-f7f. The (c) runner migrations follow in separate PRs.

🤖 Generated with Claude Code

---

Second correction: the FCI-1 hand-shell runners are class (b), not (c)

#12430 classed tools/fabric_ci_evidence_calibration.sh, tools/fabric_ci_fci1_bounded_execution_controls.sh, tools/fabric_ci_fci1_live_instrument.sh and the gunbc.fci1_bounded_execution_context_emit carrier as (c): "route available today, migration unauthored". That is wrong. The bash grammar (v2.extdeps.languages.bash rows / bash_build) has no production for constructs these runners depend on. Counts per file (calibration / controls / live instrument):

construct	calibration	controls	live
[[ ]] with glob matching	5	10	30
function definitions	2	2	6
local	0	1	25
${var: -1} substring expansion	1	0	0
$(<file) read	0	2	5
a && b || c (known serializer limit: it renders as a subshell)	1	2	7

Each header and the carrier's // annotation now name these missing capabilities as the trigger. They carry a CORRECTED 2026-09-28 line recording the (c)→(b) change, so the edit is not silent. There is no behaviour change: comments only, and bash -n is clean on all three scripts. The carrier's annotation is semantically erased (§4c).

[briansrls]

APPROVE-MERGE — content at 2f7c27c

No blocking content finding in the six-file diff. The three shell changes and the FCI-1 carrier change are comments only; the remaining delta is the correction in the rung-drop's authored prose and its matching generated documentation line. No command, workflow job, gate predicate, standing, or restoration trigger changes.

The correction distinguishes the script from the CI lane. The file-specific history includes #10390 at 2094e9c (2026-09-04) and #10270 at 9d86c98 (2026-09-05). The current FCI-1 script is real, hand-invoked shell, not proof of a restored required lane. The existing lane-restoration condition is not discharged here.

The class-(b) correction is appropriately at the grammar/serialization boundary. I inspected the bash target source and the live instrument's actual functions, local/return, double-bracket glob tests and file-read substitutions. Merely being able to emit an ordinary command does not establish representation of those control/expansion constructs. The mixed AND/OR limitation is a serializer semantic limitation, not a claim that individual AND and OR nodes are absent. No grammar expansion or broad migration is required to land this correction.

Evidence: exact-head PR workflow 36406989974 completed/success. I did not rerun bash -n, docs regeneration, or the FCI-1 live instrument; their reported local results retain that attribution. I do not independently certify every numeric entry in the per-file construct census.

Landing caveat: GitHub reports mergeable=false at this head. This is content approval, not a claim that it is currently queue-ready. Reconcile main and regenerate the projection as needed, then report/rebind the resulting SHA and use the normal queue gates. Do not restore the deleted lane as part of that reconciliation.

No merge, enqueue, live instrument execution or dispatch performed.