The simplification convergence — 2026-09-04

Lane CONVERGE. Eleven branches off main@330bbfaba, merged onto main in the map's order (docs/design/SIMPLIFICATION_MAP_2026_09_04.md §3), one regeneration, one build-and-gate pass. Every number below was measured in the convergence worktree at the commit named beside it; the recipe for each count is stated where the count is. The public cut is tools/cut.sh after every merge that was green.

1. The merge order, with each merge's conflicts and their resolutions

The rule for every conflict: read both sides; a deletion a maker made stays deleted; a file a main-lane also touched keeps both facts; a generated file is regenerated, never hand-merged. Every merge ended with cargo check --workspace --offline green before its commit.

Main moved once under the convergence (540ada0d9, the orchestrator's goal commit) and the first merge was redone on it; every later merge checked main's head before committing.

2. The six operator crates that stay, and why

crates/dclutch-svm-harness and programs/dclutch-claims-sbf/program-test/affine-batch are held on solana-program-test =4.2.1 (their manifests name the runtime panic the bump buys), and that cluster cannot resolve dclutch-operator's pins. With all nineteen absorbed, cargo metadata --offline refused the address-table interface at =3.2.0; with that dependency optional under a feature gating the twenty-two-module reach-closure of its users, it refused direct-ticket's solana-signature; with that optional too, compute-budget-interface's solana-instruction-error. A crate those workspaces link cannot live inside dclutch-operator, so the five planners they consume — market-open, market-retirement, resolution-core, product-runtime, provider-transport — stay out, and so does versioned-message, which provider-transport reaches under its own feature and which inside the operator would be a cycle. The script (tools/lane/merge-operator-crates.py) states this at its absorb list.

3. What the merge could not see, carried by hand

• trading's Dealer renames into the new accelerator's selector and into the operators branch's new carrying variants;
• formal's promotion import at the old crate root; the capability-manifest guard formal had not rewired; EmitMarketCoreRust's doc naming a crate that no longer exists;
• clients' package.json predating formal's emitter unification (five SDK scripts rewired to the one emitter per record, ts selected);
• the tier table transcribed from run.sh at the base: the accelerator's two runners, the deleted release suites, the frame gate's link count;
• the stale pins three guards carried (BAND_COUNT 25, a raw line count of 95, a compare path at the crate root) and the seventeen clippy sites in two absorbed market constituents clippy had never reached;
• three SDK generators reading what other branches deleted or moved (general-v5 → the Lean emission's V2 facts; dealer-v3 → the concept names, no selector 9; state-machines → no Dealer scenario machines);
• the browser side of two deleted units, finished: selector 9's profile mirror, fixture, vector test and the web tier's cargo-test step; the product-runtime-v2 admission generator's read of the deleted program, the SDK module's instruction half, the Studio's step 03 and the record-PDA helper only it used;
• the web's @dclutch/sdk resolves through a relative symlink: from a worktree's symlinked node_modules it reaches the LIVE tree's SDK. Every client verdict here was taken after the worktree's node_modules were rebuilt with @dclutch/sdk pointing at the converged package.

4. Restored

Nothing a maker deleted was restored. The one thing kept against a maker's intent is structural, not a restoration: the six operator crates above.

4.1 The live tree's uncommitted work

The live tree carried another lane's uncommitted work through the convergence (SERIES-5's two test files). Each fast-forward that would have overwritten a generated lock restored that lock to HEAD (every one is regenerated below); the two test files' hunks were re-applied with the crates branch's renamed paths (dclutch_capability_program_contract:: → dclutch_market::capability_program::, dclutch_release_set_contract:: → dclutch_registry::release_set::) and are dirty in the live tree again, for that lane. The untracked registered_terminal_artifacts_v4.rs moved, still untracked, from the direct-codec crate's directory to crates/dclutch-trading/src/ beside its tracked siblings; its use paths are the old crate names.

5. The line counts, before → after

One recipe on both sides, over git ls-files at the named commit: programs and crates count .rs; formal counts .lean; tools counts every tracked file minus Cargo.lock; apps and packages count .ts/.tsx/.mjs/.js; docs counts .md. The map's §0 figures (308k / 534k / 73k / 352k / 106k / 76k / 115k) are the same tree measured by hand a few hours earlier; the recipe is stated so the two columns below are one measurement.

The same seven domains under the map's own recipe — every tracked .rs/.lean/.ts/.tsx/.mjs/.py/.sh/.md file under the domain, which is what the map's §0 figures were — at 330bbfaba and at 1ef14c5a1:

docs grows because 31,782 lines of root and top-level narrative now live verbatim under docs/ledger/ (the index that replaced them is 104 lines), and because the map and this document are new. The counts that are units, not lines: programs 12 → 8; crates under crates/ 94 → 33; root workspace members 108 → 45; lockfiles 71 → 61; Lean emitters 105 → 92, the lakefile 423 → 11 lines; GOAL.md 4,992 → 104; AGENTS.md 348 → 202; routes 164 → 148, refusal codes 357 → 344, registered bands 26 → 22; the web's files 547 → 348, its twins with the SDK 202 → 3. The crates line count moves only by the deletions (a merge moves bytes, it does not remove them); the map's 0.5–0.6 M target is the rewrites' — the successor as a caller of the operator above all — which no maker started (their reports say why).

6. The ELF table

Both sets built with the release recipe — cargo build-sbf --manifest-path programs/<pkg>/Cargo.toml -- --locked — on this host (cargo-build-sbf 4.0.0, platform-tools v1.53, rustc 1.89.0), the pre-swarm set in a detached worktree at 330bbfaba (twelve links, 0 frame diagnostics), the converged set at 018ea525f (eight links, 0 frame diagnostics). No link is byte-identical: the crates branch's renames alone would have moved .strtab, and every link's .text also moved — the band macro, the split, the folds, the feature lines and the crates' layout all reach every program through the shared crates. So every link owes rows, and cohort-16 carries every redeploy (cohort-16 has not deployed). The frame rows are the recapture in §7.

The eight converged links' full digests: accelerator a8cb8c8ab35fecd70615ba7bc0451bac95d6e7d4d61547154194d9b092be8f82, claims 38196b22d71d2814452ee751cb9263ac97c75a8f7be90b695541f2e64438a7c2, core 236e92ae9345af56476509b3b89c9442fb7980c36d672aca788d4c3d1f298986, custody fd3b123a938ec376bf612025b753133b3ec97fcc52110d7dd2f889d79afc1ded, registry 98a1b45308d050378769a6d56eb1913bfe434452788d6b318c5db93e5c076d98, rent 2f145149ae575ba2180a0fa8d118dd098597620e333597bd5509cb4afe4fbcfc, resolution-proof f8bfca35ab8efaebfbce148368a1bb98ae9a58e378abadd671e5a8754b08b38f, trading 46757cedce83691cfd3fd85b1f551ca8d7d8dc7dcaab008b1ac64fa888ed4a01. The base set's full digests lived in the convergence scratchpad, which the host's fault of 2026-09-05 cleared with the rest of /private/tmp; the twelve-hex prefixes above stand, and the set is rebuilt from 330bbfaba by the same recipe. The eight converged ELFs are the release recipe's output and the cohort-16 candidate builds them again from the cut. hbox's build of the same tree (Linux x86_64, the same cargo-build-sbf 4.0.0 / platform-tools v1.53 / rustc 1.89.0) links a Trading ELF of 2753e3bf08a5…, not the laptop's 46757cedce83…: the host reaches the ELF, which is why the release builder is one named artifact and these digests are this host's.

7. The gates, with their seconds

Measured in the convergence worktree; the tree is 9f5e0aa70 (the frames baseline at 018ea525f, after which no Rust moved).

The cuts, one per green merge: ca97a0c27, 79c51f743, 6618974d8, 293c09f4c, 0f382e900, ad7fbf710, a7fc68186, abb946956, b0562e467, e287f651d, a26d5686b, 4f31eb3df, 6398fc298; the wrapper's workflow commit c21451344; then 05af32c39 and f7a834022 for the evidence and the ledger, and the wrapper's 170d913ff (the six run.sh mentions its prose kept name tools/gate). The final cut follows CONVERGE-2's commits.

7.2 The gates again at 1ef14c5a1, and what each red was

Measured by CONVERGE-2 on 2026-09-05 in a clean detached worktree at 1ef14c5a1 (the tree after the evidence and the ledger landed; no Rust moved between 9f5e0aa70 and it), warm target, one gate at a time.

7.3 The gauntlet, on hbox

Each tier at the named commit under swarm-build (MemoryMax 32 G) on hbox — Linux x86_64, cargo-build-sbf 4.0.0, platform-tools v1.53, rustc 1.89.0 — the host the Token-2022 fixture is canonical on.

The four suite reds at 1ef14c5a1, with their controls (the same runner at 330bbfaba, and at 3bee5f3f1 — after the programs merge, before the crates merge):

• core — five Series tests in found_program_test.rs (series_permit_expiry_uses_only_the_authenticated_successor_profile, series_consume_accepts_258_outcomes_and_commits_found_with_permit, series_consume_late_hoard_refusal_rolls_back_found_and_all_replay_state, series_consume_refuses_to_consume_the_same_ticket_twice, a_strangers_lamport_cannot_strand_a_scheduled_series_occurrence) refused CoreSbfError::Reference (0x3003, the Realm/Product/result-domain/Market identity linkage) at instruction 0. At 330bbfaba and at 3bee5f3f1: 15 passed, 5 failed, 2 ignored — the same five. Main's, before the swarm.
• claims-lifecycle — real_token_2022_lifecycle_refuses_ata_substitution_and_rolls_back_every_late_failure: Claims refuses 0x103004 and the lifecycle then 0x5216 (RationalLifecycleSbfErrorV2::Token) where the test asserts acceptance. No control: the runner at 330bbfaba and at 3bee5f3f1 refused --locked (that workspace's lock was stale until the one regeneration).
• dealer — real_elf_forwards_a_geometry_complete_frame_into_accelerator_authentication: 0xC006 GeneralAcceleratorSbfErrorV3::HeapFrameNotRequested where 0xC101 DealerAcceleratorSbfErrorV4::InvalidInvocation was expected — the folded link's dispatcher refuses the frame before the Dealer arm sees its content. At 330bbfaba the old Dealer accelerator's runner refused --locked; at 3bee5f3f1: the same test, the same 0xC006 for 0xC101 (1 passed, 1 failed). The red is the fold's — the programs branch — and precedes the crates merge.
• general — a_nonvacant_product_staging_cursor_refuses: ProductGraphObservation(InvalidRecord) where Product was expected. At 330bbfaba the old General accelerator's own suite: 13 of 13; at 3bee5f3f1: 13 of 13 (freeze 3, hot_instruction_v3 10), this test among them; at 1ef14c5a1 hot_instruction_v3 is 9 of 10. The one commit that touches the suite between the two is the crates merge (73499583b), and the refusal is the merged Product reader's. The red is the crates merge's.

7.1 The frames, function by function

The recaptured baseline against the pre-swarm one (330bbfaba's tools/frameguard/baseline.json: twelve links, 1,888 functions, 1,973 frames; 018ea525f's tools/gates/frames-baseline.json: eight links, 1,661 functions, 1,743 frames), every function keyed by its demangled symbol (rustfilt, the capture's <hash> placeholders substituted; no two symbols demangled alike). Of the 1,888:

• 1,322 keep their name and their frame.
• 312 are renamed with the same frame — the crates maker's claim, measured: 152 by trading's hot_v3 split into its family modules, 82 by its Dealer modules taking concept names (dealer::v3_release → dealer::release, …), 24 by the accelerator fold (dclutch_general_accelerator_sbf:: → dclutch_accelerator_sbf::general::, the Series shadow likewise), the rest by the crate merges (dclutch_market_core_codec, dclutch_rent_contract, dclutch_capability_program_contract, dclutch_execution_strategy_contract → dclutch_market; dclutch_registry_svm, dclutch_record_contract → dclutch_registry; dclutch_product_* → dclutch_product; dclutch_series_v3_kernel → dclutch_trading; dclutch_claims_svm → dclutch_claims; dclutch_source_contract → dclutch_source) and their trait impls. Eight more the matcher could not pair by name are renames too: trading's dealer::v4_{equity,lp}_accelerator_accounts::* became dealer::{equity,lp}_accelerator::* with their frames unchanged (1,856, 1,216, 64, 384, 256, 384), and v3_admitted::encode_register_bank 64 became encode_bank 64 in each.
• Exactly one grew: dclutch_claims_sbf::process_generic_plan 1,344 → 1,728, taking the dispatch that process_non_fractional_instruction (2,240 → 64) no longer does once the Core-effect route is gone.
• Five shrank: that process_non_fractional_instruction; Custody's process_instruction 3,904 → 3,840 (the reservation route gone); Trading's process_instruction 256 → 192 and series::artifacts_v3::series_base_request_digest_v3 192 → 128; the accelerator's <T as SpecFromElem>::from_elem 64 → 0.
• Four renamed and moved: hot_v3::execute_authenticated_hot_v3 3,840 → 3,776 as hot_v3::execute::execute_authenticated_hot_v3 (finding (c), the boxed stage — its 128-byte logical_projection_keys_boxed_v3 is the one genuinely new Trading function); dealer::v3_release::encode_dealer_global_program_set_v3 1,088 → 1,024 as dealer::release::…; the Dealer accelerator's process_instruction 1,024 → 64 as the fold's process_instruction (its body is now dealer::process 1,024) and evaluate_selected_family_v4 384 → 320 under dealer::.
• 209 deleted, each a unit a maker deleted whole: Custody's dealer_reservation_v1 (31) and authenticate_reservation_frame_v1; Claims' Core-effect route (9: authenticate_core_market, authenticate_core_authority, authenticate_economic_accounts, authenticate_releases, the foundational split pair, …); Trading's 161 — dealer_scenario_checkpoint_v1 (the checkpoint chain), the V2 Dealer (DealerProfileV2, interpret_projected_v2, …), dealer::v3_trade, v3_trade_artifacts, v3_composer, v3_lifecycle, v3_route, v3_admitted, the v3_accelerator_accounts scenario half, series::execute_v3's composers, series::projector, dispatch_v3's three program authenticators; the accelerator's 7 (the three constituents' entrypoints and verify_cause, two RawVec growers). And the two deleted links' 19 (direct-aot 10, product-runtime-v2 9).
• The accelerator's other new names are the fold's dispatcher (top_level_family_magic 192, dealer_family_selected 64) and the constituents' bodies re-homed with their old frames (general::process 2,176, dealer::process 1,024, series::process 64), plus general::content and dealer::content at 128 (the Dealer's was 256) and general::candidate_cause 64.

The deepest frame is unchanged: authenticate_strategy_from_sealed_boxed_v3 at 3,968 under the 4,096 bound. The diff's recipe: demangle both baselines, match exact names, then names modulo crate root, then Type::method, then a bare method name only where it is unique on both sides.

8. What is owed

Each item names its column. Nothing here is hidden behind a green.

• The gauntlet's three suite reds that are the convergence's (§7.3): dealer's geometry-complete frame refused 0xC006 before the Dealer arm — the fold's dispatcher (programs); general's nonvacant staging cursor refused ProductGraphObservation(InvalidRecord) where Product was due — the crates merge's Product reader (crates); claims-lifecycle's ATA-substitution walk refused 0x103004 then RationalLifecycleSbfErrorV2::Token where it expects acceptance — no control could run (its lock was stale at the base), so the column is claims until one does. Each is a program-test on a real ELF and none is a route cohort-16 has deployed.
• core's five Series tests refuse CoreSbfError::Reference at the base too: main's, the Series column's, before the swarm.
• tier 1: green again at cccf4d721 with the binding following its producer (§7.3); the 111 routes tier 1 never executes (37 blocked, 74 unclaimed) are the census's standing count, not this convergence's
• journey: the journey campaign's test target does not compile against the ladder (recovery_ladder, the 9-argument crank calls: RECOVERY-4's 6a3079454, 61706bc9a, 8875255a5), because the campaign's #[path] mirror of the successor was not moved with it — the recovery column; tools/fractional-exterior: compiles only once it leaves the solana-sdk =2.1.0 line (§7.2) — the drivers column.
• root-targets: the census agrees with the table (73 cheap targets, all wired — 52b84b369), and the tier then says PASS (§7.2): 73 targets, all wired, all green, after three repairs (92905bd34, e907315ae).
• clippy: the two merged crates repaired (c666a9976, b2d6660a0); the tier again at HEAD: PASS (8 s; §7.2). The four debt reds and the ten packages behind them are the standing register, each row with its owner.
• commands: the dclutch-terminal --help row, red in a fresh checkout because its dist/ is unbuilt (the gates report found the same at main) — the clients column: either the probe builds it or the runbook says so.
• Four web tests stay red, each with its owner: capabilitySelectedGate and one capabilityPhaseGate expectation — the census publishes no selected gate behind the Direct crosscheck (ROUTE_SELECTED_GATES_V1 is empty in the clients branch's census too; the gates and clients columns); explorerCoverage's record survey reads DCLTRIX1 off the SDK's protocol-constants table as a record while the explorer renders it as the Registry instruction it is (clients); tradeFlowRefusals' wall wording against the SDK's copies of the modules it reads (clients). Twelve magics the protocol-constants table declares and the explorer does not decode are exempted with the clients column named; two routes the census enumerates and the explorer does not name (the Claims conservation route, the Series arm) likewise.
• The hot-tail sweep is a probe: run-hot-cu.sh quotes M-61 only with --checked-gate and its sha256; the 20 of 20 at 44e0cf880 and its mean are diagnostics until cohort-16's release pack supplies both.
• The base ELF set's full digests were lost with the scratchpad (§6); the twelve-hex prefixes stand and the set is one cargo build-sbf at 330bbfaba from being whole again.
• The parked operator merge's six crates (§2): a program-test cluster bump (solana-program-test =4.3.0-beta.2) that does not panic would let them join; measured not to be that day.
• Rent into Core, the Series flatten, the Dealer release merge, the successor as a caller of the operator, the wire-literal census, the 45-clause refusal split — each maker's report names its own; none started here.

---

ADDENDUM, 2026-09-05, lane SUITES-2: §7.3's "11 of 15" was fifteen coin tosses

§7.3's suites row reads 11 of 15 green and §8 hangs four reds off it. That row is not a measurement of fifteen suites. It is one draw of each, and tools/gate suites at the time ran each row exactly once, so every verdict in it — green and red alike — was a single toss of whatever distribution that row has.

For at least one row the distribution was near-even. claims-lifecycle's real_token_2022_lifecycle_refuses_ata_substitution_and_rolls_back_every_late_failure passed about half its draws from d6c4dea63 (2026-08-27) onward, against one binary and one ELF: the honest RetireCoordinate is byte-identical to the nonzero-supply hostile that precedes it, so on the same recent blockhash it signed to the same signature and the bank refused it as AlreadyProcessed without running the program, and ProgramTest registers blockhashes on a wall-clock timer. Which side of that race the row landed on decided this table's cell. docs/evidence/CLAIMS_LIFECYCLE_LAYOUT_WALL_2026_09_05.md, second addendum, carries the draws.

So: the count stands as a count of one draw each, and stands for nothing more. No cell of it distinguishes "green" from "green about half the time", and §8's four reds were each also one toss — three of them were nonetheless convicted at their authors by fb36992d7 with controls at two commits, so those survive; the claims-lifecycle cell does not.

The instrument is fixed rather than the number restated. tools/gate suites now draws each row DCLUTCH_GATE_SUITE_DRAWS times (3 by default), calls a row green only if every draw is, and reports a row whose draws disagree as NONDETERMINISTIC by name — never folded into green, never into red.

Re-measured on the same host, three draws per row: 15 of 15 green, 45 runs, every one of them a pass — 16m59s for the thirteen rows that need no extra fixture, plus 72s for claims-position and claims-fractional with TOKEN_2022_SO supplied. Nothing is NONDETERMINISTIC and nothing is red.

The measured tree is a git archive export of 9fe4506f2 carrying this lane's own changes and nothing else — 5c534b016, b03b1c26e and cda85ef47's files copied in — not a clean export of cda85ef47, whose ancestry includes three other lanes' commits from the same afternoon. It is exactly the tree the row under repair was measured on before and after, and no file another lane committed that day is on any of these fifteen runners' paths; a clean export of cda85ef47 is what a cut would measure, and this is not that.