Merge pull request #1014 from AdaWorldAPI/claude/r2il-real-episodes

Real FunctionBehavior episodes: opcode matching over-admits 99.7% on real code

Author: AdaWorldAPI

.claude/board/EPIPHANIES.md

@@ -1,3 +1,58 @@
+## 2026-08-23 — E-REAL-CODE-INTERLEAVES-THE-OPCODE-MACRO-IS-NOT-A-DATAFLOW-PIPE-1 — the real FunctionBehavior episode measurement: 99.7% over-admission by opcode matching, and three structural facts the toy could not show
+
+**Status:** FINDING — [MEASURED] (`PROBE-R2IL-REAL-EPISODES-1`, 5/5, run
+against the REAL ruff-side R2IL pass-1 harvest: 143 functions from real
+x86-64 ELF64 binaries, 17,557 typed FlatFact rows, provenance FNV-pinned in
+ruff's committed `.claude/harvest/r2il/PROVENANCE.md`; bulk stream from the
+ruff GitHub Release `r2il-harvest-pass1`).
+**This closes the named-not-built item** from
+`E-HAPPY-PATH-RL-WOULD-HAVE-LEARNED-THE-CLOBBER-1` — the corpus was fetched,
+never fabricated; the probe is env-gated (`R2IL_ORE_TSV`) and exits CORPUS
+ABSENT rather than synthesize.
+**Confidence:** High for these binaries at the pass-1 seven-opcode
+convention; a 2-binary/143-function corpus is not "all real code."
+
+**Headline (E5):** of 380 real occurrences of the top recurring opcode
+trigram `(int_add, copy, store)`, exactly **1 is dataflow-chained** and 379
+are NOT — a happy-path opcode matcher over-admits **99.7%** on real code.
+Context base rate: 31.5% of ALL adjacent op pairs are def-use linked (SSA
+coverage is total — 0 of 11,653 operand rows lack a ValueId — so this is a
+dataflow fact, not missing coverage). The top idiom chains far BELOW base
+rate: it is an ADDRESSING idiom (address computed, value staged, store
+issued to memory), not a dataflow pipe. Real code interleaves independent
+chains; **sequential adjacency is not composition**. The Phase-2 R6 toy
+result (happy-path RL learns the clobber) is thus confirmed at real scale
+and sharpened: on real machine code the macro carrier must be the DEF-USE
+CHAIN, never the linear opcode window.
+
+**Refuted pre-registration, recorded (E3):** "real top-1 trigram count >
+shuffled" FAILED (380 vs 387) — shuffling a copy/int_add-dominated marginal
+CREATES monotone `(copy,copy,copy)` runs. The real structure is
+TYPE-COLLAPSE: 97 distinct trigram types vs 264 under shuffle (>2.7x), and
+the top-10 types carry 50.5% of real occurrences vs 33.6% shuffled.
+Recurring idioms are real; top-1 occurrence count was the wrong statistic.
+
+**The stamp ceiling BINDS at real scale (E4):** with 143 episodes and
+`Stamp`'s 64 bits, every widely-recurring macro measurably drops evidence —
+`(int_add,copy,store)`: 64 revised, 23 CHOICE-dropped (e=0.999). The
+modulo-64 conservatism is no longer a footnote: at real corpus size it
+discards ~26% of the evidence for the widest idiom. Conservative (never
+double-counts), but a real capacity note for Step-2's stamp residue item.
+
+**Reader validity (E1/E2):** the probe-local TSV reader is cross-validated
+against ruff's INDEPENDENTLY COMMITTED census — 17,557 rows, 5 kinds, 9
+opcodes, all counts exact; 143 episodes across 2 binaries, 5,340 ops
+partitioned. One misread caught by the gate: the TSV's kind column carries
+CamelCase variant names, not Census's snake_case (the first E1 run failed —
+the gate working).
+
+**Fences:** no `ruff_r2il` import (separate cargo workspace; schema cited,
+E1 catches misreads); reading the evidence TSV here is a measurement, not a
+re-ingest into ruff's pipeline; no mint, no BPE, no learner subsystem; the
+R2IL x BPE / OGAR-loco / V4 synthesis remains a three-IF hypothesis.
+
+**Files:** `crates/lance-graph-planner/examples/probe_r2il_real_episodes.rs`.
+
 ## 2026-08-23 — E-HAPPY-PATH-RL-WOULD-HAVE-LEARNED-THE-CLOBBER-1 — R2IL machine-state receipts are the falsification substrate the toy Phase-1 world could not provide
 
 **Status:** FINDING — [MEASURED] (`PROBE-R2IL-FRONTIER-PHASE2-1`, 7/7).

.claude/board/INTEGRATION_PLANS.md

@@ -1,3 +1,17 @@
+## 2026-08-23 — R2IL REAL-EPISODE MEASUREMENT (closes the Phase-2 named-not-built item)
+
+`PROBE-R2IL-REAL-EPISODES-1` (5/5) runs the frontier loop against the REAL
+ruff-side harvest (143 functions, 17,557 typed facts, Release
+`r2il-harvest-pass1`; env-gated, never fabricated). Measured: opcode
+matching over-admits 99.7% on the top real idiom (1/380 dataflow-chained vs
+31.5% adjacency base rate) — the macro carrier must be the def-use chain,
+never the linear opcode window; real streams collapse to 97 trigram types
+vs 264 shuffled (the top-1-count pre-registration was refuted and is
+recorded in place); the Stamp mod-64 ceiling BINDS at 143 episodes
+(~26% evidence CHOICE-dropped for the widest idiom — feeds the Step-2 stamp
+residue item). Reader cross-validated against ruff's committed census,
+exact. Entry: `E-REAL-CODE-INTERLEAVES-THE-OPCODE-MACRO-IS-NOT-A-DATAFLOW-PIPE-1`.
+
 ## 2026-08-23 — STYLE-MICROCODE FRONTIER, PHASE 2 (R2IL typed ops; the falsification intervention is load-bearing)
 
 `PROBE-R2IL-FRONTIER-PHASE2-1` (7/7) lifts the Phase-1 frontier loop