Integration 2026-09-30: #12718 #12732 #12712 #12775 #12784 #12789 #12727 #12748 #12729 #12738

[gunbai-bot]

Integration branch for ten PRs, one merge commit each, plus merges of current main (through d0fd2c6c71). An eleventh, #12523, was merged and then reverted; see below.

PRs and head shas

PR	Head	Merge
#12718 Host SHA-256 primitive sha256_hex_of_text	402742cd	clean
#12732 self-host / v2-native-cli: a non-clean emitted build exits 1 with its typed cause	b65617d2	clean
#12712 Borrow the cached module index in keyed reads	8c957d29	clean
#12523 Wall: the approval broker's import closure contains no v2 compiler module	c8b6c81a	excluded (merged clean, then reverted)
#12775 approve-ios: derive the app's server host from the broker endpoint authority	fd2e8ad1	clean
#12784 Broker front door: the backend carries its mount path	ace61c22	clean
#12789 Approval device enrolment as a handshake	37c83b6c	clean
#12727 Release-bins lookup: a tag dispatch finds its own cache entry	9745e2e9	conflict, regenerated
#12748 Generic substitution keeps a slot's optional cardinality	fdf1d075	conflict, regenerated
#12729 A service exit arm names a type	45358289	conflict, regenerated
#12738 Uniform program installation/access contract (orphaned WIP)	f2ba8742	conflict, resolved by hand; part dropped

#12523 is excluded

Its one file, test.claim.auth.approval_broker_closure_witness, is over the floor budget, so the merge is reverted here (proud-deer-538's ruling). On this PR's run 36743189391, the_approval_broker_closure_contains_no_compiler_module measured 97547 eval steps against a 72300 budget and a_closure_that_reaches_the_compiler_is_seen_to measured 12108ms against the 8000ms wall deadline; both were also refused on the enrolment margin. #12523's own head failed the same five blockers (run 36634310768), so this is not an integration effect. Moving the witness to dag/test/claim/long/ was rejected because the required floor declines that lane and the wall would stop gating merges. Head c8b6c81a stays fetchable for whoever finishes it. The annotation #12712 adds in gunbc.keyed_dependency_edge_read_seed_growth still names that witness, as one already on main does.

How each conflict was resolved

#12727 — .github/workflows/fleet-converge.yml (generated). Not hand-merged. The merge commit carries the integration side's bytes as provisional; the next commit regenerates the file from the merged authorities with tools.generated_artifact_gate main_wet_one --arg path=.github/workflows/fleet-converge.yml. The regenerated file is the integration side plus #12727's own delta (+3/-2: the second cache lookup keys on RELEASE_BINS_DISPATCH_SCOPE_REF).

#12748 — compiler_tests.rs, v1_compiler_compiler_tests_rust.rs, v1_compiler_infer.rs, v1_compiler_infer_resolve.rs (stage0 mirrors). The src/v1/*.dag authorities (04_infer, 04_method, 04_patterns, 04_resolve, 05_emit_rust, compiler_tests_rust) merged with no conflict. The mirrors carry provisional integration-side bytes in the merge commit and are regenerated afterwards.

#12729 — v1_compiler_parse.rs (stage0 mirror). src/v1/02_parse.dag merged with no conflict; the mirror is regenerated the same way.

Mirror regeneration (both of the above). claim_executor --required-regen --source-root dag --source-root src/v2, run from a seed built at the pre-#12748 integration tree. Pass 1 rewrote v1_compiler_{infer,infer_resolve,parse,compiler_tests_rust}.rs; after rebuilding from those, pass 2 rewrote compiler_tests.rs, which the seed emits from its compiled-in function and so lags one pass.

#12738 — four dag/extdeps modules and four witnesses. This branch is orphaned WIP from session/tidy-crane-517. Its first three commits already landed on main as #12630, and #12751 has since dissolved the raw argv builders it still carries.

• dag/extdeps/{bmc/webui/nbd_proxy_serve,exec/program,tools/nbdkit,tools/websocat}.dag, nbd_proxy_serve_transport_witness_test.dag, srv3/srv3_host_effect_apply_witness_test.dag: main's side. The Uniform program installation/access contract (ntfy first) #12738 side is the pre-nbd-proxy serve: raw argv builders dissolve into typed ArgvCommands at the tools' own walls #12751 shape of the same change (bmcweb_session_login_argv, nbd_proxy_serve_*_argv), and it would also have removed the websocat_nbd_proxy_bridge_command catalog row nbd-proxy serve: raw argv builders dissolve into typed ArgvCommands at the tools' own walls #12751 added. Each resolved file is byte-identical to the integration side.
• tool_pin_witness_test.dag, tool_readiness_witness_test.dag: Uniform program installation/access contract (ntfy first) #12738's side. Both sides already have the tool_invocation helper; Uniform program installation/access contract (ntfy first) #12738 adds only the WORKAROUND annotation on it.
• Kept: the new row gunbc.recurring_failure_mode nested_field_read_through_a_binder_of_an_inferred_generic_result that the annotations cite. I re-ran its reduced module on the merged seed (which includes Generic substitution keeps a slot's optional cardinality; VariantNotFound located at the pattern #12748's infer changes): the two-level read through the inferred-generic binder still fails at evaluation with error type cascade, and the one-level and annotated-scrutinee controls pass. The row's claim holds on this branch.
• Dropped: the WIP commit f2ba8742, which adds only dag/zz_scratch_typecheck/{repro,evalrepro}.dag. They are scratch probes with no consumer (§3c), and repro.dag calls compile_dag_diagnostic_census without importing it. Nothing of the "ntfy first" contract named in the PR title exists in the branch beyond what CliTool carries its ProgramIdentity (program-install lane PR 2) #12630 landed, so there is nothing further of Uniform program installation/access contract (ntfy first) #12738 to integrate.

Checks run locally

• claim_executor --required-regen and --required-regen-fixed-point at 087be8e835; the later commits (the Wall: the approval broker's import closure contains no v2 compiler module #12523 revert and a second main merge) touch no src/v1 file: first_generation_equal=true planned=161 executed=161 adjudicated=161, and fixed_point_equal=true referenced_first_generation_equal=true.
• tools.docs_projection_gate regen: exit 0, no committed file changed.
• cargo clippy --all-targets -- -D warnings: clean.
• cargo fmt --all --check: clean.

Do not merge from here; proud-deer-538 lands it and closes the source PRs.

🤖 Generated with Claude Code

[gunbai-bot]

PR #12805 floor is red with 5 blockers, all on test.claim.auth.approval_broker_closure_witness, the witness #12523 adds. It is not an integration effect: #12523's own floor at c8b6c81 (run 36634310768) was red with the identical 5 blockers, so that head was never green. Measured on #12805 run 36743189391: the_approval_broker_closure_contains_no_compiler_module = 97547 eval steps against a new-witness budget of 72300; a_closure_that_reaches_the_compiler_is_seen_to = 12108ms against the 8000ms wall deadline; both also enrolment_measured_over_margin. Both claims pass their verdicts; the cost is the live import_closure_live walk itself, so I see no small edit that brings a 12s walk inside the margin. The other two import_closure_live witnesses live in dag/test/claim/long/, which the required floor declines, so moving it there makes CI green but the wall stops gating merges, which is the stated purpose of #12523. Its author wrote that an over-budget result would come to you as a cost finding rather than a debt row. Everything else on the floor passed (planned=621 passed=607, the rest held rosters); generated is green; emit-build still running. A = revert the #12523 merge on the integration branch (10 PRs land; #12523 returns separately once its cost is solved). B = keep it, move the witness to the long lane.

— sent from wise-bee-807

[gunbai-bot]

On review 73317's note that #12738 has no content left: it does have content, 25 lines, all in git diff origin/main...HEAD:

• the new row gunbc.recurring_failure_mode nested_field_read_through_a_binder_of_an_inferred_generic_result (+17);
• the WORKAROUND annotation on tool_invocation in tool_pin_witness_test.dag and tool_readiness_witness_test.dag (+4 each).

The rest of #12738 was already on main as #12630/#12751, and its scratch probes were dropped. The PR body describes both.

The enrolment-handshake files (gunbc.auth.approval_enrolment_handshake and its witness) come from #12789 (commit 3df5fe5), not #12738, and are merged as that PR's head 37c83b6c. No change pushed.

— sent from wise-bee-807