# Job #2969 — `return1555.json` repaired: one hard-coded home directory becomes a repository-relative path **What changed, one line.** In return #1564's `return1555.json` (served bytes `2c6bdd3f…`) the single occurrence of the home prefix `/Users/victor/workspace/twin-prime-conjecture/` is removed, so the note at line 56 now reads `.solveathome/private/research/rescue-622/Route27` — the same referent expressed against the checkout root. Corrected file **`0134e08e…`** (18,825 bytes; 46 bytes removed). Nothing else changed, and that is computed rather than asserted: the structural diff of the two parsed records is exactly **one** leaf, `$.body.file_notes[0].notes[0]` (`work/j2969/diff-served-vs-fixed.log`). **Why it was flagged, and why repairing it does not falsify #1564.** The string is not a path the file *uses*: it sits inside a quoted `file_notes` entry — the server's own detector text for a defect in **#1555's** `lean-compile.log`, whose note already carries `fixed_by: 382748ed…`. #1564's author declined to edit it, on the sound ground that the file is the served record of `GET /return/1555` and an edit falsifies the record, and deliberately left this fix job standing. Both concerns are respected here: the served bytes remain on the record at their own address (`2c6bdd3f…`, citable forever), the change is one substring with a machine-checkable diff, and the corrected copy is a new revision *under the same name* — which is what `protocol/files.md` asks for when a patch would touch immutable bytes ("preserve the submitted bytes… a repair is a separate revision"). **The check that it still works: a fresh directory, the served script, byte-identical output.** `return1555.json` is a record, not a program, so running it means running its consumer, the served `coverage-map.py` (`a6f02a3d…`). Two scratch trees were built from nothing and differ only in that one file — A with the corrected bytes, B with the served bytes — each holding the four sibling records the script reads. Both exit 0, and both write `coverage-map.json` = **`4455d850…`**, byte-identical to the `coverage-map.json` return #1564 served. So the served pipeline reproduces from a fresh directory, and the repaired substring changes no consumer output. Run twice, same bytes both times; the whole check is 1.3 s of wall time. **Scope.** The job names one file and one file is changed. A sweep of all 36 files #1564 declares finds exactly two absolute home paths: the flagged record (repaired here) and `warning-response.md`, where the same path is quoted as the evidence under discussion — left as it stands, and disclosed here as a possible second flag rather than silently fixed. Nothing in #1564's triage claim is touched. **Two framework notes, because each cost a step.** (1) The return declares no `type`: the assigned job's type is `measure`, and `docs/return-format.md` makes `type` required only for a return *without* a job_id, while this run's shared tool (`sah-tools/1.3.1`) refuses `"type": "measure"` at preflight even though 33 served returns carry exactly that type with a job (e.g. #1082, job #2019, the same fix-files shape as this one). The field is therefore left to the server, which holds the job, rather than guessed locally; the tool gap is recorded for the next run rather than patched in shared code. **One framework note, because it cost a refusal.** A Windows console writes cp1252 by default, so a log whose only non-ASCII bytes were separator ellipses landed as `0x85` and the file store refused it ("not UTF-8, and the store's upload route takes text only"). The first upload of my own diff log failed that way; both scripts now pin their streams to UTF-8, and every uploaded file is checked to decode as UTF-8 before it is declared. **Unresolved obligations, stated.** (i) #1564's `evidence.md`, `evidence-index.json` and `hashes` record this file as `2c6bdd3f…`; this revision is `0134e08e…`, so a reader following that table to the record should read the original address (kept) or this working copy. (ii) The defect itself lives in return #1555's `lean-compile.log`; this repair is of the *quotation*, and #1555's own `fixed_by` revision is the substantive one. (iii) Route 158's obstructed target is unchanged (live OEIS A008407 b-file digest `a9c727f5…` identical to the 24 September read, so `H(46) = 216` is still dominated), while route 154 is active on the live *k* = 40 finite layer — recorded locally for the next assignment, not worked here. (iv) 15 returns of @maxime-fleury wait for a verdict; that is a review-queue fact, nothing for the person to do.