CI floor time: child-spawn attribution (Pi-3 bench × srv1) — corroborate + correct #7106

[briansrls]

Summary

Measurement-only follow-on to the merged #7106 CI-floor audit. Same mandate, re-run on a
Raspberry Pi 3 used as a slowdown-exaggeration bench against srv1 (Ampere Altra, same
class as the CI runners), with release binaries built from the merged head. Everything is by
execution; no floor behavior, substrate, or unit-modeling changes — the only .dag edit is one
doc-graph bind row.

Corroborates #7106: phase walls, resolves_total=4, materialization receipt, width=1 at the
16 GiB cgroup ceiling.

Corrects the attribution:

• The band's largest single class is cold child processes, which CI floor time audit: redundant-work ledger + ranked levers (45–72min band) #7106 does not name. Every
  wet gate discharges witnesses through run_gunbc_claims_ready (dag/tools/host_prelude.dag:177),
  which spawns one cold gunbc process per ClaimRun. 26 children across batches 1/6/7 ≈
  23 min ≈ 47.5% of the 48.4-min floor — invisible to resolves_total=4, identical on trivial
  diffs. It's the third instance of "cold index per execution unit" (per-thread Floor time: dissolve the #6848 once-per-entry fixpoint + reconcile-miss qualified-fill (~18min residual) #7030, per-repeat
  session/double-resolve-rewire, per-process here).
• Measured counterfactual: pooling batch-1's 12 claims into one process is 479.6s → 93.4s
  (5.1×) on srv. But on the Pi, pooling ≈ spawn-sum — the win vanishes at the memory wall
  (the unioned closure swap-thrashes), so the fix must be paired with per-worker footprint
  reduction or it won't help the capped fleet runners.
• CI floor time audit: redundant-work ledger + ranked levers (45–72min band) #7106's ranked levers, corrected: lever Add SVG viz, test helpers, and makegen scaffold #1 (per-entry fixpoint) is stale — Floor time: dissolve the #6848 once-per-entry fixpoint + reconcile-miss qualified-fill (~18min residual) #7030/floor: precompute both-closure edges to dissolve per-entry bare-ref fixpoint #7056
  are ancestors of the measured run's commit with no recovery; lever Codex/graph viz test helpers #2 (ingest re-walk) targets
  the wrong mechanism — the discover_source_root_ingest bin is 0.008s (2 file reads), the
  12 min is the child-spawn class.
• Two unledgered floor costs: post-floor selection control (4m51s/PR) and retained-store
  teardown (~2.5m, swap 5.5→9.4 GB during Drop).
• Side finding (needs root-cause, not a timing claim): standalone gunbc compile --target dag
  over the three witness-layer roots reds with 2652 UnlistedImportUse errors on the CI-green
  tree; the floor gate consumes the executor's internal shared-index receipt instead — two
  compile-clean realizations, opposite policy on the class the namespace lane is mid-promotion on.

Recommendations (plan doc §8): dissolve the child-spawn class toward the process_shared_index
"resolve once, share by reference" terminal (interim: pool per gate), gated on footprint;
re-open the discovery 643s as an unknown; ledger the two unledgered costs; keep a footprint probe
on the emit stage (337× Pi stretch — the most memory-shaped compiler stage).

Carriers

• docs/plans/ci-floor-child-spawn-attribution.md — narrative, corrections, recommendations
• docs/probes/ci_floor_pi_srv_stretch_2026-07-23.tsv — phase × CI × srv × Pi × stretch × shape
• docs/probes/ci_floor_child_spawn_counterfactual_2026-07-23.tsv — pooled-vs-spawned ledger
• dag/gunbc/doc_graph_roots.dag — one HandAuthoredDocBind row → tools.host_prelude::run_gunbc_claims_ready

Test plan

• doc_graph_has_no_orphan_docs / doc_graph_has_no_dangling_links / slug wall — verified by execution on srv1 (results in-thread)
• No ci_floor_plan batch ordering or floor runtime changes (diff is docs + one doc-graph bind row)
• Bind target tools.host_prelude::run_gunbc_claims_ready resolves in-tree
• CI floor (measurement docs only; fleet gate)

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.