P1 retention vs drain cohort receipt (authoritative)

dag/gunbc/doc_graph_roots.dag

@@ -398,6 +398,13 @@ data hand_authored_doc_binds: List<HandAuthoredDocBind> = [
     additional_works: [],
     dissolution: HasTrigger { text: "Nat lives only in std.nat (coproduct + Peano ops), v2.std.nat imports rather than redefines it, the 138-entry co-occurrence class is gone, and integer's GroupCompletion<Nat> cites the unified type — then this per-concept census doc deletes with this row (full criteria in the doc §9)" },
   },
+  HandAuthoredDocBind {
+    home: PlanDoc,
+    slug: "p1-retention-vs-drain-cohort-receipt",
+    primary_work: DeclarationRef { module_path: "gunbc.p1_retention_cohort_receipt", decl_name: "p1_retention_cohort_receipt_reproduction_steps", field: WholeDeclaration },
+    additional_works: [DeclarationRef { module_path: "gunbc.p1_retention_cohort_receipt", decl_name: "p1_retention_cohort_receipt_verdict", field: WholeDeclaration }],
+    dissolution: HasTrigger { text: "the redirected assembly/materialization lane this receipt's REJECT verdict points to is picked up and either supersedes this pinned measurement with a canonical carrier or the parent floor-prep-tax-program plan closes P1 outright — then this receipt and its anchor module delete together, per the same scaffold trigger named beside p1_cohort_receipt_enabled in cli_run.rs." },
+  },
   HandAuthoredDocBind {
     home: PlanDoc,
     slug: "push-event-affected-set-baseline-design",

dag/gunbc/p1_retention_cohort_receipt.dag

@@ -0,0 +1,7 @@
+module gunbc.p1_retention_cohort_receipt
+
+import std.types { String }
+
+data p1_retention_cohort_receipt_reproduction_steps: String = "git checkout the commit where p1_cohort_probe.rs first lands; ctrl-build --local -- cargo build -p v1-compiler --bin p1_cohort_probe --release; run twice over the fixed 50-entry cohort in p1_cohort_roster.txt — Mode A with GUNBC_FLOOR_DRAIN_RETENTION=1 GUNBC_P1_COHORT_RECEIPT=1 (default eviction ON), Mode B additionally with GUNBC_SCHEDULE_RETENTION_EVICT=0 (retain-all pole); compare [p1-cohort] wall_ms/resolve_ms per entry over the valid overlapping range. Full receipt: docs/plans/p1-retention-vs-drain-cohort-receipt.md."
+
+data p1_retention_cohort_receipt_verdict: String = "REJECT — entries 2..34 (the valid Mode-B overlap before its OOM kill at entry 34/50) do not stop repaying the ~1-2s per-entry tax when schedule-derived retention eviction is disabled; mean/median wall_ms and resolve_ms are statistically indistinguishable between Mode A (eviction on) and Mode B (retain-all). Schedule-derived retention/eviction is not the cause of the tax; redirect to assembly/materialization reuse."

docs/plans/p1-retention-vs-drain-cohort-receipt.md

@@ -0,0 +1,124 @@
+# Receipt — P1 retention vs drain cohort receipt (authoritative)
+
+**Status:** measurement receipt, timestamped 2026-08-03. **DESIGN.md + `docs/plans/floor-prep-tax-program.md` (PR #7721) remain authority** — this is a dated receipt, not a fact ledger.
+
+**Lane:** floor prep-tax program, P1 · **Subject:** does schedule-derived retention (#7129 eviction) cause the ~1s per-entry "tax" observed across the floor discovery path, or does the tax persist independent of eviction? · **Deliverable:** measurement only, ACCEPT/REJECT verdict against the parent-stated criterion. No retention-mechanism change lands in this PR.
+
+**Main SHA measured:** `dddc79596a9dbb9bcbe4ac1c77121a302adb6e77` — the commit where `p1_cohort_probe.rs`/`p1_cohort_roster.txt` first land in git history (the probe/instrumentation source was authored and built against this same tree state before that commit was made; the earlier `b53aec1d` — this PR's initial WIP commit, predating the probe — cannot reproduce the runs below, since `p1_cohort_probe.rs` does not exist in its tree; correction per review 47754). Reproduce from `dddc79596a9` or any later commit on this branch.
+
+**Vehicle:** new `p1_cohort_probe` binary (`src/v1/stage0/src/bin/p1_cohort_probe.rs`) — a thin CLI front end that calls the **same production entrypoint** `run_discovery_corpus_with_options` that `claim_executor` drives for the real floor path, with `DiscoveryWidthPolicy::Adaptive` (the governor-admitted width the fleet actually runs under — schedule-retention only arms inside this branch's width-1 special case), over an explicit fixed 50-entry cohort instead of a full-corpus scan. This is not a reimplemented/toy mechanism; it is the identical discovery-corpus/schedule-retention code `claim_executor` uses, scoped for direct entry-by-entry A/B comparability. `claim_batch` was investigated and ruled out as a vehicle: it does not call `run_discovery_corpus_with_options` at all (a different assembly-cost-decomposition harness — see `docs/plans/per-entry-assembly-decomposition-measurement.md`).
+
+**Cohort:** 50 entries, `src/v1/stage0/src/bin/p1_cohort_roster.txt`, sourced verbatim from `docs/plans/receipts/entry-graph-union-slice2/receipt-post-merge-representative-50.json` (`summary.entry_timings[].entry`) — the same representative-50 module-sharing cohort an earlier slice-2 receipt already established as sharing the module universe.
+
+**Host regime (honesty):** dashboard session container, uncapped-relative-to-fleet cgroup (see cgroup readback below); this is a regime-relative measurement, not a byte-match to the fleet width-1 receipt in `docs/plans/m2-floor-retention-measurement-receipt.md`.
+
+---
+
+## 0 — Reproduction (verbatim invocations)
+
+```bash
+git rev-parse HEAD   # dddc79596a9dbb9bcbe4ac1c77121a302adb6e77 (or later on this branch)
+CTRL_BUILD_MODE=local RUSTC_WRAPPER= CARGO_BUILD_JOBS=4 \
+  ctrl-build --local -- cargo build -p v1-compiler --bin p1_cohort_probe --release
+cp -f target/release/p1_cohort_probe target/release/p1_cohort_probe-snapshot
+# sha256: 54566635522137569193c1e56451ccc1a38c6e4e5595cc6fca57aee2cc26943a
+```
+
+**Mode A — default eviction ON (production default):**
+
+```bash
+export GUNBC_FLOOR_DRAIN_RETENTION=1
+export GUNBC_P1_COHORT_RECEIPT=1
+LOG=docs/probes/p1_retention_cohort_20260803T022255Z/mode_a_eviction_on.log
+target/release/p1_cohort_probe-snapshot > "$LOG" 2>&1
+```
+
+**Mode B — retain-all measurement pole:**
+
+```bash
+export GUNBC_FLOOR_DRAIN_RETENTION=1
+export GUNBC_P1_COHORT_RECEIPT=1
+export GUNBC_SCHEDULE_RETENTION_EVICT=0
+LOG=docs/probes/p1_retention_cohort_20260803T022255Z/mode_b_retain_all.log
+target/release/p1_cohort_probe-snapshot > "$LOG" 2>&1
+```
+
+Raw logs and extracted `[p1-cohort]` lines are checked in under `docs/probes/p1_retention_cohort_20260803T022255Z/`.
+
+---
+
+## 1 — Mechanism receipts (both modes armed correctly)
+
+Mode A:
+```
+[floor-drain] schedule-retention ARMED: entries=50 modules_refcounted=711 evict_enabled=true
+```
+
+Mode B:
+```
+[floor-drain] SCHEDULE-RETENTION EVICTION DISABLED (GUNBC_SCHEDULE_RETENTION_EVICT=0): arming + counting only, dropping NOTHING — the retain-all measurement pole (pre-M2 process-lifetime retention). schedule_evictions will be 0 BY CONSTRUCTION; unset the env to restore eviction.
+[floor-drain] schedule-retention ARMED: entries=50 modules_refcounted=711 evict_enabled=false
+```
+
+Both modes armed over the identical 50-entry/711-module schedule. Mode B's loud diagnostic fired as required and `modules_evicted=0 graphs_evicted=0` held on every one of its 34 completed `[p1-cohort]` lines — confirms Mode B genuinely never evicted (not a partial/soft eviction).
+
+**Field-naming note (post-run correction):** the raw logs checked in under `docs/probes/p1_retention_cohort_20260803T022255Z/` were captured with the original field name `schedule_cache_hit`, which review 47754 correctly flagged as mislabeled — the value logged is `typecheck_compute_count()` movement (a typecheck-memo hit/miss), not schedule-retention cache occupancy. The field is renamed to `typecheck_cache_hit` in `cli_run.rs` going forward; the raw logs are left unedited (they are the actual captured evidence) and every value/analysis in this receipt reads from that same underlying signal under its corrected name — no data changed, only the label.
+
+---
+
+## 2 — Completeness (honest accounting, not fabricated)
+
+**Mode A completed all 50 entries** (414 witness rows, 1 pre-existing unrelated failure — `witness_observed_hostname_reads_typed_op_hermetic`, an environment-specific hostname-observation test, not a retention/timing concern; identical failure occurs in Mode B at the same entry, confirming it is unrelated to eviction mode).
+
+**Mode B was killed by the OOM killer at entry 34/50** (`exit=137` / SIGKILL). Its `process_rss` climbed monotonically and unboundedly across every entry — 1.72 GB at entry 1 to 5.33 GB at entry 34, `cgroup_memory_current` reaching 8.07 GB — because retain-all, by construction, never frees a module. This is itself a first-class, expected finding of the retain-all pole (unbounded process-lifetime retention is exactly what M2's schedule-derived eviction exists to prevent), not an instrumentation bug. **Entries 35-50 have no Mode B data and none is fabricated or interpolated here.** All comparison below is restricted to the valid overlapping range, entries 1-34.
+
+---
+
+## 3 — Quantitative A/B comparison (entries 2..34, the valid overlapping range)
+
+| | Mode A (eviction ON) | Mode B (retain-all) |
+|---|---|---|
+| mean `wall_ms` | 2045.4 | 1936.7 |
+| mean `resolve_ms` | 1943.9 | 1891.6 |
+| median `resolve_ms` | 1076 | 1054 |
+| min `resolve_ms` | 869 | 908 |
+
+Per-entry, the two modes track each other almost exactly (full table computed from the raw logs, entries 1-34):
+
+```
+entry  A_wall  B_wall  A_resolve  B_resolve  B_rss_MB
+    1    2409    2247       2372       2236      1716
+    2    1155    1024       1076       1019      1716
+    3    1175    1009       1090       1001      1716
+   10    1768    1750       1728       1735      2025
+   11   13131   12397      12843      12186      2681   (heavy entry, both modes)
+   18    3669    3393       3594       3363      3241
+   32   11368   10737      10378       9911      5034   (heavy entry, both modes)
+   34    1140     913       1070        908      5085
+```
+
+(Full 34-row table and extraction script output retained in the probe log directory.)
+
+The heavy entries (11, 18, 21, 27, 29, 32) are heavy in **both** modes at nearly the same magnitude — eviction is not what is driving their cost, and retain-all does not relieve them either. The steady-state ~1-1.1s-per-entry baseline (entries 2-9, 12-17, 19-20, etc.) is likewise statistically indistinguishable between A and B; if anything Mode B trends marginally faster (noise-level, consistent with slightly fewer cache-eviction/re-resolve cycles, not a collapse).
+
+**No entry in 2..34 shows the tax collapsing toward near-zero under Mode B.** Retaining every module for the full cohort life (growing to 5+ GB RSS) does not make entries 2..34 cheap.
+
+---
+
+## 4 — Verdict
+
+**Parent's stated ACCEPT criterion:** *"ACCEPT iff entries 2..N sharing the module universe stop repaying ≈2s tax under Mode B."*
+
+**REJECT.** Entries 2..34 do not stop repaying the tax under Mode B — mean/median `wall_ms` and `resolve_ms` are statistically indistinguishable between Mode A (eviction on) and Mode B (retain-all, arming confirmed, zero evictions confirmed, OOM-killed only after unbounded growth at entry 34). Disabling schedule-retention eviction entirely does not collapse or materially reduce the per-entry tax on this cohort.
+
+Per the parent's own routing instruction, this receipt reports the result the A/B data shows, not a pre-decided answer (the earlier `claim_batch` framing that was retracted mid-task is not relied on here — this conclusion is derived solely from the floor-path A/B receipt above). **The ~1s-2s per-entry tax is not caused by schedule-derived module eviction/retention.** The redirect is to assembly/materialization reuse (as the parent's fallback routing anticipated, but arrived at here by measurement, not assumption) — consistent with the open thread in DESIGN.md's "duplicate work = the content-hash qualification of `Materialization`" and the per-entry assembly decomposition receipts already in `docs/plans/`.
+
+**No retention mechanism change is implemented in this PR** — per the explicit brief, this receipt's job was discrimination, not a fix.
+
+---
+
+## 5 — Artifacts
+
+- `src/v1/stage0/src/bin/p1_cohort_probe.rs`, `p1_cohort_roster.txt` — the probe vehicle. **Scaffold, not a permanent second floor driver:** P1's REJECT verdict is now recorded, so per the dissolution trigger named alongside `p1_cohort_receipt_enabled` in `cli_run.rs`, this probe and its instrumentation delete once no other open lane needs cohort-scoped A/B retention receipts — not kept in tree indefinitely.
+- `src/v1/stage0/src/cli_run.rs` — `[p1-cohort]` per-entry instrumentation, gated independently by `GUNBC_P1_COHORT_RECEIPT=1` (kept separate from the existing `GUNBC_FLOOR_DRAIN_RETENTION` switch — two distinct facts, two distinct toggles, DESIGN §3). The emitted field is `typecheck_cache_hit` (whether `typecheck_compute_count()` moved during the group — a typecheck-memo hit/miss signal), not a schedule-retention cache-occupancy read (schedule-retention has no per-group hit/miss concept, only the cumulative `modules_evicted`/`graphs_evicted` counters already carried on the same line); named honestly after review 47754 caught the earlier `schedule_cache_hit` label as mismatched to what it measured.
+- `docs/probes/p1_retention_cohort_20260803T022255Z/` — raw Mode A/B logs and extracted `[p1-cohort]` line sets.