phase-36: delever_regen also starts from an agent's lever-free near-miss (e22: the families finished its 28 body in one pass); e22 banked 3 (ov_SC06_029); func_80185D44 stays at its one lever

docs/SETUP.md

@@ -1819,7 +1819,7 @@ CLAIM, not a fact — two were refuted on bytes on 2026-09-10.**
   ~29 s once). Each judged class is appended to **`.run/P36/regen/<L>.jsonl` as it lands** — the agent lane draws only
   classes the pass has judged and not closed. An ALL-families pass (`--families R2 … R26`, `-j 26`) costs ~1.8 s per `--try`
   and 100–370 candidates per head class; it closed an R12 class in its first minute, so re-running the old families on a
-  moved tree is not redundant.
+  moved tree is not redundant. **S104 (e22):** a third start text — a pack's lever-free `body.c` (an agent's near-miss); the sweep had only seen the free body, and the families finish some agent bodies in one pass.
 - **`delever_search.lever_free_body` refuses a strip that unbalances block comments (S104, R43)** — a lever line that OPENS a `/* …` comment continuing onto the next line was deleted whole, leaving the comment's tail as code; the start text never compiled and every sweep read the class as UNSCORED. Known-true: e24's pre-bank func_80180E24 now refuses; negative control over the whole residue: 709 ok, 16 other refusals, 7 comment refusals — each checked to be a real `/* …` lever line.
 - **`tools/alloc_table.py` prints the INTEGER priority (S104)** — `global.c:594-603` stores it in a `register int` (truncated) and breaks ties by allocno number (`:607`); equal integers are flagged `TIE`. The float column hid e12/e14's closes (245.6 vs 245.2). Known-true: e14's `func_8017BEBC` lever-free dumps show the two `&g.sz` pseudos at 245 TIE.
 - **`delever_pack.related_bodies` for main (S104):** main's units are `src/800*.c`, not `src/main/` — the search now falls back to the target TU's directory; every main pack's `related.txt` had been empty (e3: the answers were same-TU siblings sharing a global). Known-true: `func_80036FB0`'s pack now lists 7 bodies sharing `D_8006AEE8`/`D_80078F10`.

tools/delever_regen.py

@@ -61,6 +61,13 @@ def starts(e):
     bp = ds.RUN / "bodies" / f"{e['alias']}__{e['fn']}.c"
     if bp.exists():
         out.append(("best", bp.read_text(errors="surrogateescape")))
+    # an AGENT's lever-free near-miss (S104 e22, func_8017C954: the sweep had only ever seen the 632 free body; re-running
+    # the families on the agent's improved 28 body found the finisher in one pass — 1,655 candidates, two at 0)
+    ap = REPO / ".run" / "P36" / "agents" / f"{e['alias']}__{e['fn']}" / "body.c"
+    if ap.exists():
+        t = ap.read_text(errors="surrogateescape")
+        if not re.search(r"__asm__|\bregister\b[^;]*\$", t) and all(t != x for _n, x in out):
+            out.append(("agent", t))
     return out