Files
BFM-decomp/docs
Drew T a2da99b130 src(phase-36): T6 rung D banked — the four exemplars the permuter closed, tidied and byte-judged (R22 218/218)
- delever_permute --bank: 4 winners applied through delever --apply-body, each IDENTICAL on its own object; the winner is
  now TIDIED first (pycparser reprints a body it parsed: two-space indent and a corpse `;` where a statement was inlined
  away) and the tidy is judged like any other candidate, so the source keeps the tree's shape; parenthesisation and brace
  style are deliberately left to the formatting phase, over the whole tree at once
- a slip, named: --bank re-applied a permuter body over the ONE-LINE version rung R had already banked for
  func_80135D20; the clean text was restored through the oracle (label d1fix) and --bank now skips any body the ledger
  already calls LEVER-FREE
- --recipes is killable now: the oracle writes the candidate into the tree to compile it, so the original goes into
  inflight.json first (P35's rule: a tool restores from its own snapshot) — a killed run had left a candidate in src/
- rung R's R6 generalised from "assigned once, read once" to "dead after one read" (the lever rung D actually found:
  uVar5 is assigned in two branches and only one was inlined); it still does not reproduce that class's win, which is
  recorded as an open item rather than papered over
- lever_census: 33,953 sites (was 33,957), 0 UNMARKED; lever_progress snapshot "T6 d1"; R22 218 passed, 0 failed of 218
2026-09-09 12:21:42 -06:00
..