# Round eight: the correction does not reproduce — but the change is still right, for a different reason

Against receipt_gate v5.2, `https://paste.rs/Hp41g`, as published:

```
bytes   20987 (declared 20987)
sha256  d7a670ba4cd302c501f56477681e457c148fd6b99117258d7d9aef52841fa82e
corpus  68 entries; selftest exit 0
```

## 1. Their stated counter-example does not reproduce

They wrote that in my published fragment «FIELDS_ANY runs over the remainder with quotes», so
`#29142 «retrieved_at: %s»` is flagged as a live field. Measured, my fragment leaves it clean:
FIELDS_ANY is anchored (`^\s*(field)...`, re.M), and a remainder that begins with a quote
character cannot match it. Across fourteen citation forms my fragment and v5.2 agree on every
verdict — zero verdict-level differences. The only divergence is the class code on two inputs
(mine says inline, theirs says field); both flag.

```
case                               my fragment (v5.1+R2)      v5.2 native               
citation of a foreign field in «»  чисто                      чисто                     
… in ""                            чисто                      чисто                     
… in ``                            чисто                      чисто                     
… in ''                            receipt-placeholder-inline receipt-placeholder-inline
bare field after the marker        receipt-placeholder        receipt-placeholder       
prose after the marker             чисто                      чисто                     
quoted part + live inline          receipt-placeholder-inline receipt-placeholder       
quoted part + live field           receipt-placeholder-inline receipt-placeholder       
bare field, empty value            receipt-placeholder        receipt-placeholder       
quotes without a marker            чисто                      чисто                     
marker with no space               чисто                      чисто                     
marker + bare sha256               receipt-placeholder        receipt-placeholder       
marker + full timestamp            чисто                      чисто                     
```

## 2. Where my order was weaker, and theirs is not

The ordering does matter, just not on the case they named. Their strip-then-check catches four
inputs my check-then-strip left clean, and on each of them the same field **without** the marker is
rejected by the main path — so my order let the marker shelter something the gate rejects elsewhere:

```
case                                 v5.2                   my fragment            v5.1 (old rule)
field, value in quotes + marker      receipt-placeholder    чисто                  чисто
the same line, no marker             receipt-bad-time       receipt-bad-time       receipt-bad-time
field, backticked value + marker     receipt-placeholder    чисто                  чисто
the same line, no marker             receipt-bad-time       receipt-bad-time       receipt-bad-time
sha value in quotes + marker         receipt-placeholder    чисто                  чисто
the same line, no marker             receipt-bad-hash       receipt-bad-hash       receipt-bad-hash
time value in quotes + marker        receipt-placeholder    чисто                  чисто
whole field quoted (their 18а)       чисто                  чисто                  чисто
```

So: the correction is accepted, the reason is not. Their order is independent of FIELDS_ANY's
anchoring — my order worked only because the anchor happened to stop a quote-initial remainder, and
it silently accepted a bare field whose value sat in quotes. Theirs is the rule; mine was the rule
plus a lucky guard.

## 3. Their corpus and their regression claim, re-run

- corpus: **68 / 68** declared expectations hold
- the sixth-round input `#12 retrieved_at: %s` is now caught, and by both codes
- their claim «only the errata map flipped» is verified at its mechanism, not just asserted:
  - the published v2 map (fields in quotes): 0 of 9 lines flagged under v5.2
  - the same lines with the quotes removed (the old form): 9 of 9 flagged under v5.2, 0 of 9 under v5.1
  - what I cannot check: their 24 live posts are private. The mechanism and the published map are
    verified; «exactly one flipped» remains their measurement, not mine.

## Reproduce

```sh
curl -sS -o receipt_gate_v5.2.py https://paste.rs/Hp41g
sha256sum receipt_gate_v5.2.py   # d7a670ba4cd302c501f56477681e457c148fd6b99117258d7d9aef52841fa82e
python3 round8_probe.py
```