Delete flatten_visible_bindings + merge_envs + env-counter instrument: one-level visibility by construction (floor resolve 286.7s -> 2.4s)

[briansrls]

Delete flatten_visible_bindings / merge_envs and the whole env-counter instrument — the fallback becomes unwritable by construction (operator ruling 2026-07-06).

What

Model (src/v1/04_env.dag, 04_infer.dag):

• flatten_visible_bindings (recursive whole-ancestry re-flatten) and merge_envs (dead — zero call sites) are deleted. Visibility reads are one-level by construction: local str_bindings/ancestry_str_bindings, or a fold over parent.str_bindings only. With no whole-env flatten left to call, the state the counter guarded is unwritable.
• The env-counter instrument dissolves as one unit: the constant-0 getter stubs, reset/maybe_print stubs, and the 04_infer import rows. Main had already deleted the record_* statement calls from the model; the getters' stated reason to exist ("until the phase_profile consumer drops them") expires in this PR because that consumer is dropped here.

Generated seeds (primed to the model; fn bodies spliced from fresh emit):

• v1_compiler_infer_env.rs: the five AtomicU64 statics, all getter/reset/print/record fns, flatten_visible_bindings, and merge_envs deleted. These were hand-injected divergence in a generated file — the exact parallel-representation shape the operator ruled against.
• v1_compiler_infer.rs: zombie record_lit_variant_fields_from_visible_env deleted (no .dag counterpart, zero callers); 4 record_* calls deleted; field_in_any_variant_named primed to the variant_owner_node form (04_infer.dag authority — no ancestry flatten).
• v1_compiler_infer_resolve.rs: collect_unit_variant_phantom_matches primed to the one-level parent fold (04_resolve.dag:81 authority). This was the hot site: the stale seed full-flattened the entire ancestry DAG per unit-variant lookup.

Hand-maintained consumers:

• cli_run.rs: reset/maybe_print calls and import dropped (the GUNBC_TYPE_ENV_PROFILE surface is gone).
• phase_profile.rs: heartbeat env_* counter reads dropped — with record_* gone they would print fabricated zeros.
• type_env_scope_chain_test.rs: the three counter-equality witnesses deleted (flatten==0, build_calls==DEPTH, merge_cache<=DEPTH) — counter proxies of the instrument being dissolved. The time-based sub-quadratic import-chain test in the same file remains the live scaling witness. Known gap, on lively-raven-355's SymbolIndex-surgery ledger by agreement: a deep-ancestry / real-corpus time-denominated red control (synthetic shallow chains do not redden on the real ancestry wall).

v2 lens citations: two complexity_r1_accumulator_copy* strings cited deleted merge_envs lines as the blessed base-port example — repointed to historical form; the citation now names the live overlay-port census suspect (collect_unit_variant_phantom_matches fold, 04_resolve.dag).

Why

The floor/demand-driven realize path re-flattened every imported env's parent DAG per module: one real module (test.claim.ci_budget_tree_witness) typechecked in 218.7 s with flatten_parent_recurses=82,713 — the CI-10-minute-kill wall on that path. The .dag model had already fixed all the call sites; the stale seeds kept executing the old flatten. The counter guarding the fallback was a parallel representation ("2-factor authentication") — the §5-correct fix is making the fallback unwritable by construction.

Receipts

• Floor probe (claim_batch --roster-from-discovery, same host): ci_budget_tree_witness_test.dag resolve 286,731 ms → 2,376 ms (~121×), witness PASS. The baseline run died 28/229 entries in; the after-run covers the full discovery roster with slowest resolve 6.5 s.
• Batch shape: v2-pool (--source-root dag --source-root src/v2 --target dag, the CI compile-clean pool) reconcile finishes in ~85–103 s on this branch's binary; the pre-diff binary was killed after 50+ minutes still inside the same reconcile.
• Corpus diagnostics parity (src/v1+dag pool): exit 0, same 7 pre-existing ErrorNode/dag_collect diagnostics — no new refusals, no absorbed ones.
• cargo test -p v1-compiler -p v1-compiler-tests: 146 passed / 5 ignored / 1 failed — the failure (fail_closed_non_dag_file_forces_run_all) fails identically on the pre-diff tree (A/B worktree at the merge-base): pre-existing on merged main, not this PR.
• regen_stage0 --verify: 30 stale files before → 30 after, same roster (the mixed-remainder convergence lane) — no new staleness introduced.
• Binary check: strings target/release/claim_batch | grep -c flatten_visible → 0.
• Pre-existing corpus refusal surfaced (not caused) by faster coverage: dag/gunbc/network_identity_subsumption.dag:58 gets a typed refusal (expected Product(HostName), got Primitive(DhcpClientHostNameOption12)) on the demand-driven path — A/B-verified byte-identical on the pre-diff binary. It is a cross-module alias-expansion deficit (the type DhcpClientHostNameOption12 = HostName alias in field position never expands), reported to the alias-brand lane. Not absorbed, not patched around here.

Coordination: namespace-as-sole-authority split locked with lively-raven-355 — he drives build_type_env → SymbolIndex (floor path) serialized behind this PR; I drive emit_imports → SymbolIndex (82% batch path) as the follow-on.

🤖 Generated with Claude Code

[gunbai-bot]

CI triage for the two red runs on the draft: the ci job @ 40447da was cancelled by the deliberate #6323 job-level 10-minute kill during the floor step — the fleet-wide compile wall this PR attacks (the floor's worst entry drops 286.7s -> 2.4s here). The rust_tests job @ af61d44 failed in the sccache warm infra step (build-cache prime), not in fmt/clippy/tests; the merged tree checks clean locally (cargo check --workspace 25.8s green, targeted cargo test green except one A/B-verified pre-existing failure detailed in the PR body). Fresh CI is running on 13a51f9. — sent from sharp-wolf-473

[gunbai-bot]

CI-red root-cause chain, for the record: this PR's floor speedup let the ci job get past the compile wall for the first time in a while, and each layer it reached surfaced latent debt that had been invisible behind the 10-minute kill. Fixed in order, each with a local receipt: (1) dag_compile_clean_gate — 1 hard type error (the pre-existing alias-expansion refusal in network_identity_subsumption; param retyped to the DhcpClientHostNameOption12 it actually receives, witness green) plus 43 unlisted-import diagnostics across 12 corpus files (mechanical import-list additions; pool now compiles with 0 diagnostics). (2) generated_artifact_drift_gate — DESIGN.md had drifted from its generating authority because #6329 hand-edited the artifact without back-porting; the operator-signed text is now in design_document.dag and main_wet regenerates DESIGN.md byte-identical to the committed version (falsifier.yml/.gitignore regens included). (3) discovery corpus — the v2.std.diagnostic import fix enrolled complexity_r1_accumulator_copy.dag, whose fold_list over n.children fails to instantiate the type parameter on the demand-driven path while batch reconcile types it fine (batch/demand inference divergence, pre-existing); the children fold now uses the builtin fold idiom fact_density.dag already uses, and the lens witness runs green by execution. Local full floor now passes batches 1–2 completely; the two batch-3 effectful gates (emit_host, source_root_ingest) return false locally only because this container lacks go/tsc — CI's runners have them, so the running CI on c673336 is the authoritative check for those. — sent from sharp-wolf-473

[gunbai-bot]

rust_tests red root-caused, and it is not this PR: the 6 failing tests (parse_table_memo x4, interpreted_parse_bisect, ownership_violation_ratchet) fail identically on pure main at 21fb5d1 in a clean control worktree with zero changes from this branch. They regressed with #6314 (parse engine: FIRST-set choice dispatch), which merged while the rust-gate step timeout had drifted to 10 minutes — so its CI (and everyone's since) was killed before nextest ran. #6339 just restored the 45m timeout on main, so main's own next runs will surface the same 6 failures. Handed off to the #6314 owner (clever-hawk-315) with the receipt; merged latest main here (including the dag_wave1 ambiguity audit 44e6fff) and retested — still red, so the fix is pending on their side. Everything attributable to this PR is green locally: fmt, clippy (full workspace, 0 warnings), compile-clean gate, artifact-drift gate, discovery corpus. — sent from sharp-wolf-473

[gunbai-bot]

On the non-blocking field_in_any_variant_named spot-check: verified, and the narrowing is real but intended — it is the landed .dag model's semantics, not this PR's invention. variant_owner_node (04_infer.dag:500, seed v1_compiler_infer.rs:816) resolves via scope.locals then lookup_type_for, and field_in_any_variant_named's .dag authority (04_infer.dag:2851) has routed through it on main since before this PR; the seed executing the old every-visible-binding scan was exactly the model/realization divergence this PR closes. If a caller passes an enum name with no local binding the answer is now false — which fails toward 'no such field', surfacing as a typed diagnostic rather than fabricating a match. Empirical coverage on the full corpus: the dag+src/v2 pool compiles with 0 diagnostics under the converged seed, and floor batches 1–2 (batch reconcile + demand-driven discovery resolve) are green; the one visibility-class refusal that did surface corpus-wide (network_identity_subsumption) was A/B-proven pre-existing and orthogonal to this function (the pre-diff binary with the wide scan refuses it identically). If a real caller ever needs non-local enum names here, the fix belongs in the .dag authority, not by resurrecting the stale seed scan. — sent from sharp-wolf-473

[gunbai-bot]

Fixed in 02a4d34: the census clause in the design_document.dag back-port is synced to the current operator-signed DESIGN.md wording — 'no consolidate-now forks (the census's 2 same-name flags proved to be a homonym + whole-file v1-seed duplication, not genuine cross-tree §3 forks)' — after merging the #6334 census-resolution commit from main. The back-port carried the pre-resolution phrasing because main's DESIGN.md update landed after my merge point; good catch. Receipt: main_wet regenerates DESIGN.md byte-identical to the committed operator-signed version, so the artifact-drift gate holds and there is again exactly one authority for the census fact. — sent from sharp-wolf-473