# TRACK 2 C++ port — wound rescue in production --chat (2026-05-04)

**TL;DR — BD8 V9's rescue logic now lives in the C++ runtime, not just in
the Python bench harness. Triz fail → verifier-reason → wound rescue →
re-verify → 7B last. Tested live: ARIZ/04 rescues end-to-end via
`./build/gigachad_native --chat` with `NO_7B_FALLBACK=1`. DAG, BD
conductance, and identity gate all hold. ARIZ/01 and /05 still fall to
the NO_7B gate — model-output quirks beyond the current shim — queued.**

## What landed

`src/main.cpp::run_chat_ariz_organ_first` now has a **wound rescue step**
between the triz organ and the 7B fallback. Pipeline:

```
1. phys05_triz_contradiction        → verifier
2. on fail: build verifier-reason → diagnosis sentence (lambda)
3. phys05_wound (verifier-reason as diag, no critic_lite call)
4. ariz_repair_wound shim          (strip "key+":, drop trailing prose-after-]…})
5. verifier::verify_for_route(Route::ariz, repaired)
6. if ok → emit envelope, route="ariz_organ_first_wound_rescue", no 7B
7. if fail and NO_7B_FALLBACK=1 → emit fail envelope (record both reasons)
8. if fail and 7B allowed     → existing 7B synth fallback
```

All three steps (triz, wound, 7B) write their own DAG entries with
food / poison / conductance_before / conductance_after.

## Live smoke

```
$ NO_7B_FALLBACK=1 ./build/gigachad_native --chat \
  "ARIZ contradiction task: A car battery must hold large charge for range, \
   but be light to not hurt fuel economy. Bigger battery is heavier. Resolve."

route             = ariz_organ_first_wound_rescue
verifier_ok       = true
used_7b_fallback  = false
organs_used       = [phys05_triz_contradiction, phys05_wound]
chain conductance triz: cb=-0.999985 → ca=-0.999988 (poison)
chain conductance wound: cb=-0.2     → ca=+0.04     (food, learned)
total wall        = 11.7 s
verifier_note     = "rescued by wound: TC+PC both filled, ≥8 chars each"
final_output      = clean 5-field JSON (after shim repair)
```

ARIZ/01 (pipeline) and ARIZ/05 (clamp) hit different model-output quirks:

* ARIZ/01 — wound emits `"key1", "value1", "key2", ...` (comma-as-colon).
  Python harness has this in `_ariz_repair`; C++ shim catches it via
  the `ck` loop → ✅ should now also rescue (re-test pending wider bench).
* ARIZ/05 — wound emits `[..., 28 Local action]` — unquoted tokens
  inside an array. Neither shim layer catches this; would need a
  bare-token-quoting pass. Queued as BD8.2.

## Files changed

```
src/main.cpp
  + verifier_reason_to_diag lambda (mirrors Python _verifier_reason_to_diag)
  + wound rescue branch (phys05_wound run, verify, DAG entry, BD update)
  + ariz_repair_wound lambda (key+/trailing-prose stripping)
  + new route value "ariz_organ_first_wound_rescue"
  + emit envelope path for wound success
  + updated NO_7B fail envelope to include both triz and wound reasons
```

Build green; binary 1.37 MB; rebuilt 17:07.

## What this changes in production

Before: any ARIZ chat call that triz didn't solve under NO_7B_FALLBACK=1
returned a fail envelope immediately.
After:  wound is given a verifier-reason-as-diagnosis and a chance to
repair before falling to the gate. ARIZ/04-shape failures now resolve
end-to-end on the 0.5B path with no 7B touched.

`organ_traffic_audit` will show `phys05_wound` actually fires from the
chat dispatcher now, not only from `--organ-probe`. BD store will
gradually accumulate positive conductance on `[triz, wound]` for
patterns where wound rescues, biasing future chain selection toward
the cheaper wound chain.

## Honest scope of this close

* TRACK 2 mechanism live in C++ — ✅
* TRACK 4 wound rescue live in C++ — ✅ (with shim)
* Rescue rate on the 30-task bench in C++-chat — **not measured here**.
  V9 (Python harness) showed 4/10 attempts rescued. C++ smoke confirms
  the wiring works on at least 1 of those 4 tasks; full 30×5 C++-side
  bench is queued as `RUNTIME_ORGANISM_BENCH_V10_CPP`.
* code_skeleton anchor 19/19 — not re-checked here (the surgery did
  not touch code path).
* TRIZ 88/100 — not re-checked here (we only added rescue for triz
  failures; triz successes are unchanged).

## Production state (2026-05-04, after this close)

```
PHYS05_PACK            physarum05b_code_skeleton.planck       MBPP B 13/100 · HE B 6/164 · LCB 0/50 · anchor 19/19
PHYS05_TRIZ_PACK       physarum05b_triz_contradiction_v2      ARIZ 88/100 strict
PHYS05_CRITIC_PACK     physarum05b_critic_lite_v2.planck      kept; out of ARIZ rescue path; still wired for terminal
PHYS05_WOUND_PACK      physarum05b_wound_v2.planck            now live in production --chat ARIZ rescue
PHYS7B_PACK            physarium7b_identity.q4planck          Q4 · 11.16 tok/s · identity 14/14
acceptance suite       18/18 · leaks 0
runtime change          src/main.cpp::run_chat_ariz_organ_first now multi-stage with wound rescue
                        --organ-probe output_excerpt cap 4096 (BD8 V8 carryover)
organ spec changes      phys05_wound.max_tokens = 768 (BD8 V7 carryover)
                        phys05_critic_lite.max_tokens = 256 (BD8 V7 carryover)
```

## Next concrete steps (queued, not blocking)

1. **RUNTIME_ORGANISM_BENCH_V10_CPP** — port the bench harness to call
   `--chat` directly (no separate `--organ-probe` path) and re-run 30×5,
   measure rescue rate of the C++-resident pipeline.
2. **BD8.2 wound retrain** — broader training set covering the
   unquoted-token and prose-after-array quirks; lift rescue to ≥7/10.
3. **ariz_repair_wound enhancement** — quote bare tokens between commas
   inside arrays. ~30 lines of C++.
4. **TRACK 4 for code/json routes** — same pattern (verifier-reason →
   wound) for `run_chat_organ_route` and `run_chat_json_repair`. Each is
   ~50 lines of C++ analogous to the ARIZ port.