Determinism: construction compile-gate in, x2 sampling gate out (Phase D)

[briansrls]

Determinism: construction gate in, x2 sampling gate out (Phases C.1-live + D)

Completes the determinism-mechanism plan's gate-replacement: make v2.lens.determinism the standing forward guard over the v2 compile door, then delete the x2 sampling gate it replaces.

Why (§5 — construction over validation)

The old tools.emit_determinism_gate was validation: compile the corpus to Rust twice and diff -r (~12m, corpus-denominated). That proves "two runs coincided" — a sample, not "output is a function of declared inputs". C.2 (#6576) made the v1 seed emitter map_keys-free, so emit determinism is now a construction property. Removing a post-hoc check that construction superseded is exactly §5's "prefer the single authority from which the realization is derived over a check that flags it after the fact."

What lands

Live construction gate (completes C.1's AC4):

• v2.lens.determinism.determinism_compile_gate — a rooted whole-tree findings walk (one located diagnostic per nondeterministic-root call node), enrolled in v2.compiler.compile.always_required_root_lenses, mirroring accumulator_copy_compile_gate (Complexity as a compile citizen: accumulator-copy enrolled as a required root-grain compile lens #6527). Reds the instant a bare map_keys reach enters the v2 compile door; vacuously green today (v2 has zero map_keys calls).
• Discriminating controls in reach_witness_test.dag: RED on a planted map_keys call and a nested-in-subtree reach; GREEN on a clean call and the sorted_map_keys deterministic sibling.

x2 gate deleted (Phase D / AC6):

• Removed EmitDeterminismGate from type Gate — compile-red-guided the full consumer touch-list across ci_gate, ci_gates, floor_effect_gate_witness, floor_materialization, merge_admission, commit_workflow (incl. the whole GithubActionsEmitDeterminismJob CI-job surface), ci_spec, ci_workflow, ci_floor_plan.
• Deleted dag/tools/emit_determinism_gate.dag + dag/tools/emit_determinism_transport.dag; regenerated ci.yml (the emit_determinism job is gone — build → ci is now the whole serial chain).
• Re-homed the duplicate-computation lens witness (computation_demand_duplication_witness_test.dag) onto a self-contained synthetic ReplicatedOracle fixture (it had borrowed the x2 transport as its TypedArgvInvocation example; its property is independent of emit-determinism).
• heavy_pair_spec scheduler control re-pinned to RegenVerifyGate (heavy + host-compiler + off-floor: the faithful analog of the removed gate, preserving the exact heavy count).

Coverage / scope (honest boundary)

The gate is enforced exactly like accumulator_copy (#6527): via floor-enrolled fixture tests and every validate_then_compile door invocation (self-host witnesses) — growing as v2 self-hosts. DagCompileCleanGate drives the v1 seed, so this is the forward wall for the v2 door, not a whole-corpus live scan. v1 emit stays construction-only (map_keys-free by C.2) with no active guard — operator ruling 2026-07-14 (b): a dying §7 seed carries no active guard; the construction guarantee replaces the sampling oracle. Phase B (physics-grounding via kind: Observation) and Phase F (cross-function derivation) remain gated/deferred.

Verification (green-by-execution, on the merged tree)

• Full compile-clean over dag + src/v2 (the DagCompileCleanGate): 0 diagnostics.
• Generated-artifact drift gate: ExitSuccess (ci.yml == authority).
• 9 witness tests green: the 4 determinism gate discriminators; witness_gate_roster_matches_coproduct_arms (roster == type Gate arms — the consistency guard for the arm+roster removal); affected_set_universe_gate_processes_match_declared_gates; witness_plan_serializes_heavy_resolves (scheduler serialize-heavy with the RegenVerify substitute); the two re-homed duplicate-computation controls.
• Re-verified after merging origin/main (which touched 00_compile.dag/fix(compile): subtree required-lens roster O(n²) cost-shape (fact_density lane) #6579, commit_workflow.dag, ci_floor_plan.dag and the v1 seed) and rebuilding the seed binary.

🤖 Generated with Claude Code

[gunbai-bot]

The CI red here is the pre-existing main-red, not this PR.

The only failing check is ci at batch-1 dag_compile_clean_gate_passes (returned Bool(false)) — the always-whole-tree compile-clean gate (fn_arrow_decl_substrate_is_whole_tree false, blocked-on-#6239). Reproduced locally on this head (6a85823):

non-exhaustive match: missing variant(s) Srv3InstallReconcileObserve, DurableApprovalGrantRecord,
DurableApprovalGrantRead, Srv3InstallMediaFetch, Srv3SeededInstallMediaRemaster, ... Srv3ReceiptEmit
(dag/gunbc/fleet_converge_cli.dag :: converge_cli_applied_knob_count)

Cause: the shell→dag srv3/converge slices (#6572/#6586/#6598) expanded the host-effect coproduct with ~12 Srv3* / DurableApprovalGrant* variants; the match in fleet_converge_cli.dag — a file this PR does not touch — wasn't updated. Because the whole-tree compile-clean gate is always-on and stops batch-1 before everything else, this reds every open PR (11 of 12 open PRs red; only #6604 green).

Owned and already fixed by #6604 (adds _ => fallback to converge_cli_applied_knob_count; CI-green, mergeable). I'm not pushing a duplicate fix into a load-bearing lane I don't own (a wrong fallback over an effect coproduct is a §5 masking risk). Once #6604 merges and green main is re-merged here, this PR's compile-clean gate clears; I'll re-verify the full floor (batch-2 witnesses + the determinism_compile_gate) on the re-merged tree before calling it ready. Phase D itself was CI-green standalone at 113e7f41c (pre-merge base, whole-tree compile-clean receipt ok=true).

— sent from deep-koi-605