audit: math-fix pass 34 post-push addendum (f20e62a, 34/34 sweep re-run vs new origin/main)

MATH-FIX-REPORT.md

@@ -935,3 +935,5 @@ Determinism: `passes3-pass34.txt` (`01075d0e…`, 21 consecutive), `after-couple
 Harness notes: Playwright ran with `TMPDIR=/home/ubuntu/projects/retire-planner/.pwtmp/tmp` (harness TMPDIR read-only; no site impact); `scripts/deploy.sh`, `audit/mathfix_capture.py` (canonical leg) and `audit/offlan_check.py` run unmodified from the repo; ssh byte-check uses `-n`; sweep34 rebuilt under `.tmp/mathfix34/` (gitignored) from the pass-33 original with tag retarget only (`diff`-proven: header + 2 list paths; header ordinal fixed `34rd`→`34th`), lists regenerated from `git ls-files` + `find -printf '%P\n'`/`comm` (`diff`-proven byte-identical vs pass-33: 21 tracked + 13 ignored); all three mixed probes copied byte-identical (md5-proven) and run with argv `after34` (the probes read TAG from `sys.argv`, not env); FIRE/mobile legs via the md5-identical prefix-tagged copy of `mathfix_capture.py` (artifacts moved into `audit/math-fix/` — the copy resolves `OUT` relative to itself); engine-side mixed probes inline in node with app-normalized rates (`app.js inputs()` divides percents by 100 — first naive attempt at fraction-as-integer rates reproduced the 350%-growth signature and was discarded, not recorded). First probe launch passed TAG via env → probe fell back to its default tag `after23` and re-drew the gitignored `after23-mixed-rows.json`/`-table.png` — re-verified **byte-identical to the lineage** (`96665bbf…`/`222c81a1…`), zero evidence impact. Rowsweep result note: the pair heuristic on the mixed JSONs reports the documented **(53,54)** (reachable) and **(55,56)** (52/56) growing-Total-with-Withdraw pairs — both rows carry ` †` + the legend (`daggerRows: ["54"]`/`["56"]`), i.e. exactly the plan's intended honest footnoted mixed window ("total grows — fine"); the probes' per-row validators (canonical for mixed JSONs) are green, and the canonical/non-mixed set prints `ROWSWEEP-OK`. Image inspection of the PNGs is byte-verified against the visually confirmed renders of passes 1–33 (identical md5 + IHDR where stable; frame variants enumerated above); this harness model is text-only (`read_image` not used — same policy as passes 24–33); PNG pixel diagnostics via Pillow (numpy absent).
 
 No code change required — plan executed end-to-end; fix commit `08398267040a832bdec0910626423f0dbfeb0a7f` ("engine: GIS 65+ floor, clawback only with OAS, couple mixed-row footnote, FIRE retire age 40") is at `origin/main` (confirmed ancestor of `HEAD` via `git merge-base --is-ancestor`; working tree byte-identical to it under `site/`), so this pass re-ran every acceptance gate fresh against the deployed build. Final: **MATH_FIX_DONE**.
+
+Addendum (post-push, same pass): audit commit `f20e62a29c62a45948b3b1f705755846d003aa0c` pushed `b8bef71..f20e62a` → `origin/main` (report + fresh off-LAN stamp `2026-10-09T23:01:30Z`; code untouched — the plan's mandated fix commit `08398267040a832bdec0910626423f0dbfeb0a7f` remains the code change under test, confirmed ancestor of `HEAD`). Post-push `git fetch` + full byte sweep re-run vs the new `origin/main`: `git diff b8bef71..f20e62a -- site` = **0 lines** (zero site delta, so the sweep carries over by construction), origin blob == local == webroot md5, tracked **21/21 (3-way)** + gitignored `site/vision/*.jpeg` **13/13 (2-way)** = **34/34 match, 0 mismatch** → `SWEEP-OK`; `HEAD` == `origin/main` == `f20e62a`.