Regen round cost: claim_executor --regen-round-cost prices one convergence round on two clocks through gunbc.regen_round_cost, and the affected-set disposition

[briansrls]

Regen round cost: a named phase-breakdown instrument + affected-set scoping

Deliverable 1 — the instrument. claim_executor --regen-round-cost --source-root dag --source-root src/v2 runs ONE convergence round exactly as gunbc.generated_artifact_merge_driver generated_artifact_merge_driver_repair_steps prescribes it — seed build → the same run_required_regen the required phase runs → install of the drifted mirrors → rebuild from the installed seed → diff — and prints every phase on two clocks (wall, process-tree CPU), rendered by gunbc.regen_round_cost regen_round_cost_render through the interpreter (no Rust copy of the format). It extends the existing instrument rather than adding one: v1_rt::trace_mark (emitted from v1.compiler.runtime_rust) gains a CPU reading per mark and an armed ledger; the regen host stamps its own phase boundaries under the labels gunbc.regen_round_cost regen_round_phase_label names; compile.dag's existing marks are consumed unchanged. Receipt also lands at target/stage0-regen-round-cost.txt.

Fail-closed arms: the emitting seed must be the built seed (executable bytes digested before/after the seed build; a changed binary refuses with the reason); a refused regen stops the round at emit with no install; a failed rebuild-from-installed is carried on the receipt as a round failure (the phases before it are measured facts) and fails the exit code; an unreadable CPU clock renders na, never 0, and poisons the total.

Evidence: dag/test/claim/regen_round_cost_witness_test.dag (9 fns incl. two RED controls, all true on srv1 via the built gunbc), and two Rust lockstep tests in required_regen_host.rs (host_built_receipt_renders_through_the_model is byte-exact against the model; run remotely, green).

Two receipts, one edit (planted data row in std.content_hash, tree 677988a2, clean):

BuildBuddy runner 192.168.241.2 (6 cores, 7 GiB):

regen-round-cost: producer=claim_executor --regen-round-cost host=192.168.241.2 tree=677988a26ddb4d8c73ac3351045d8c17aa5ea660 tree_dirty=false seed_build_compiled_crates=1 rebuild_compiled_crates=1
regen-round-cost: phase=seed_build wall_ms=227832 cpu_ms=382180
regen-round-cost: phase=corpus_load wall_ms=36072 cpu_ms=36060
regen-round-cost: phase=corpus_reload wall_ms=17315 cpu_ms=17310
regen-round-cost: phase=compile.frontend wall_ms=9188 cpu_ms=9190
regen-round-cost: phase=compile.normalize wall_ms=392 cpu_ms=390
regen-round-cost: phase=compile.reconcile wall_ms=23222 cpu_ms=23220
regen-round-cost: phase=compile.analyses wall_ms=431 cpu_ms=430
regen-round-cost: phase=compile.emit wall_ms=158899 cpu_ms=158890
regen-round-cost: phase=mirror_write wall_ms=17198 cpu_ms=17480
regen-round-cost: phase=candidate_verify wall_ms=0 cpu_ms=0
regen-round-cost: phase=adjudicate wall_ms=17003 cpu_ms=17290
regen-round-cost: phase=hand_verify wall_ms=82 cpu_ms=80
regen-round-cost: phase=digest wall_ms=32961 cpu_ms=33530
regen-round-cost: phase=install wall_ms=0 cpu_ms=0
regen-round-cost: phase=rebuild_from_installed wall_ms=238766 cpu_ms=399410
regen-round-cost: phase=diff wall_ms=5 cpu_ms=0
regen-round-cost: total wall_ms=779366 cpu_ms=1095460
regen-round-cost: changed_paths=1 [src/v1/stage0/src/std_content_hash.rs]
regen-round-cost: FAIL regen: generated surface drift: std_content_hash.rs

srv1 (128 cores, load ~50 during the run):

regen-round-cost: producer=claim_executor --regen-round-cost host=srv1 tree=677988a26ddb4d8c73ac3351045d8c17aa5ea660 tree_dirty=false seed_build_compiled_crates=1 rebuild_compiled_crates=1
regen-round-cost: phase=seed_build wall_ms=373448 cpu_ms=621040
regen-round-cost: phase=corpus_load wall_ms=60478 cpu_ms=60270
regen-round-cost: phase=corpus_reload wall_ms=28377 cpu_ms=28280
regen-round-cost: phase=compile.frontend wall_ms=20599 cpu_ms=20550
regen-round-cost: phase=compile.normalize wall_ms=544 cpu_ms=540
regen-round-cost: phase=compile.reconcile wall_ms=38458 cpu_ms=38340
regen-round-cost: phase=compile.analyses wall_ms=634 cpu_ms=630
regen-round-cost: phase=compile.emit wall_ms=295857 cpu_ms=294910
regen-round-cost: phase=mirror_write wall_ms=24119 cpu_ms=24470
regen-round-cost: phase=candidate_verify wall_ms=0 cpu_ms=0
regen-round-cost: phase=adjudicate wall_ms=23892 cpu_ms=24250
regen-round-cost: phase=hand_verify wall_ms=121 cpu_ms=130
regen-round-cost: phase=digest wall_ms=46407 cpu_ms=47250
regen-round-cost: phase=install wall_ms=0 cpu_ms=0
regen-round-cost: phase=rebuild_from_installed wall_ms=366445 cpu_ms=617300
regen-round-cost: phase=diff wall_ms=15 cpu_ms=10
regen-round-cost: total wall_ms=1279394 cpu_ms=1777970
regen-round-cost: changed_paths=1 [src/v1/stage0/src/std_content_hash.rs]

Fixed-point control (no edit, tree 75bc5007, srv1): total 480.7 s wall / 480.9 s CPU, changed_paths=0, first_generation_equal=true.

What the instrument says (shape, not transcribed as a claim beyond these trees):

• The regen process is single-threaded end to end (CPU ≈ wall in every regen phase). Emit is the largest regen phase (≈51 % of the regen process on both hosts) but only ≈20 % of a round with a real edit.
• The two cargo builds of the seed are ≈60 % of a real round (BuildBuddy 228 s + 239 s; srv1 under load 395 s + 366 s) for a ONE-mirror change, because every mirror compiles into one crate. The change-denominated cost lives in the rebuild, not the emit — so the addendum's D3 (shard the emit) condition (“emit is the dominant share”) is NOT met and D3 was not built.
• Cost-shape defects the receipt makes visible: the corpus is read twice (corpus_load then corpus_reload inside compile_stage0, 53 s/77 s), and rustfmt normalizes the whole population three times (mirror_write, adjudicate, digest: 67 s/97 s). Both are fixable and are not fixed here (measurement lane).
• The two-round bootstrap is measured, not assumed: round 1 on both hosts (tree 0fe2c517) had the stale seed emit the OLD v1_rt.rs over the hand-synced one and the rebuild refused; round 2 (tree 75bc5007) reached the fixed point.

Deliverable 2 — affected-set disposition (gunbc.regen_round_cost regen_affected_set_disposition): DerivableNow from the closure the regen already builds (cli_run regen_input_sources_over_roots / extend_sources_to_both_closure_fixpoint: imports + dotted + bare refs), read in reverse from the edited module and restricted to the compared mirrors, plus the runtime template products. Evidence by identity: (1) edit v1.compiler.runtime_rust → drift {v1_compiler_runtime_rust.rs, v1_rt.rs} vs 9 predicted (v1_rt.rs is the one dependent outside the import graph: the template's product, emitted from the string baked into the emitting seed); (2) edit std.content_hash → drift {std_content_hash.rs} vs 72 predicted, on both hosts. The bound is an over-approximation; the residual is the seed's own baked mirrors as a generation-level input (the bootstrap), and exactness needs the per-module emit. The cross-module state named in the model (reference_derived_use_lines closure-wide args) stays inside the reverse closure, so it does not break the bound.

Also fixed on the way: resolved_graph_cache::transform_content_digest read current_exe(), which becomes … (deleted) after any in-process rebuild re-links the binary; it now reads /proc/self/exe (the running image) on Linux. Observed on both hosts: the in-round cargo build recompiled the v1-compiler crate once more after an identical build in the same shell (seed_build_compiled_crates=1, byte-identical output) — src/v1/stage0/build.rs declares rerun-if-changed=src; attributable, not chased here.

Not built, cited for follow-up: gunbc.witness_floor_workflow runs with cache: false deliberately (rust-cache's post step deleted other jobs' files on the shared self-hosted home; the note beside that row says ~/.cargo persists across jobs anyway). A cache row there is a candidate only for target/, and the receipts show the seed build is the dominant cost of a round, so it is worth pricing separately.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KGCnzm2zkb37mVHgrGCt9L

[gunbai-bot]

Addressed review 57566 in fce29f5: TraceMark no longer carries wall_ms: Int / CpuMs { ms: Int }. Each mark now carries durations: List<std.observation.TimedMeasurement> — one row per clock (std.measure.ClockBasis WallClock / CpuClock) whose elapsed is Measured<std.measure.Nanosecond>, so an unreadable CPU clock is MeasuredUnavailable { cause } on the model's own carrier and no unit lives in a field name. Totals fold through observation_duration_of per basis; the receipt renders milliseconds as a projection of the nanosecond value (regen_round_ms_text). The host builds the same carriers (Measure { count } / MeasuredValue / TimedMeasurement) and the byte-exact lockstep unit test plus the ten .dag witnesses (one added: an unreadable CPU row does not poison the wall total) are green on the reworked model.