# vendor_row33_p2_twins -- the row-33 P2 higher-order-twins data shelf: the served evaluator's fixed-eps values at dps 140 at the further point P2 = (s, t, m^2) = (-3/2, -1/4, 1) against a pair of independent AMFlow grids at auxiliary-mass orders 400 and 500, the compare receipt with the per-eps members, counts and planted-digit controls, the two grids' leg receipts, the eight evaluator receipts, the statement objects and the timing line of record (row 33, LBL3VP; 2026-09-10)

Release note (2026-09-29): the compute-host labels in the vendored receipts, statements and price line (values and key names alike) were re-cut to neutral letters (hostA / hostB), and the grid receipts' `stamp_fields.cgroup` value to its unit name; the sha256 given for each served object below is the bytes after that re-cut, the `as emitted` sha256 is unchanged.

DATA only.  No served script reads this directory: `lbl3vp-evaluate.py` (the reference-point evaluator; its dps-110 default stands), `lbl3vp-points-evaluate.py` (the four-point evaluator; default dps 110) and `lbl3vp-eps0-evaluate.py` are unchanged, byte for byte, and pin nothing here.  The eight dps-140 values on this shelf were produced by the same evaluator build the served `vendor_row33_points/eval_row33.py` is the vendored copy of (module sha256 `8c804aab1409ba7876127cfaf02bdf1521602e9292d6473d628f147e8ffe078e`, the receipts' `module_sha256.eval_row33.py`; the served copy differs in docstrings, comments and message strings only, as the bundle's changelog of 2026-09-04 states), run per eps at `--dps 140` with the quadrature levels 7 / 8 (the receipts' `PRODUCER.argv`) -- the recorded setting; the served default stays dps 110.  This shelf carries the objects behind the printed clause of 2026-09-10 (the two printed strings below): at P2 the dps-140 value agrees with the x500 grid to 76.33-77.87 digits over the eight eps (the counts, int(min) per eps, 76 76 76 76 77 77 77 77: the cell figure 76, the range 76-77), the twins agree with each other to 118.41 digits at worst (120.41 at best), the evaluator's own L7/L8 pairs read 79.45-80.58, and the dps-110 / dps-140 pair 37.58-38.83 = the served floor (the served page's 37-38 figures).

THE FIGURES (read from `compare/COMPARE_ROW33PP_P2_20260910T045045Z.json`; the rule is its `rule` field: `agree(a,b) = -log10(|a-b|/|b|) at mp.dps 400 on the full strings (gate_point.py agree()); counts int(min), never round(); a verdict at n_compared 0 is VACUOUS`; convention A = the reference-point rule, count = int(min) over {vs the x500 grid, the L7/L8 pair, the x400-vs-x500 agreement}; convention B = int(min) over {the dps-110-vs-dps-140 pair, vs the x500 grid}; the bar 50 digits; verdict `PASS at >= 50 d at all eight eps (convention A); the planted controls FAIL by name`):

| eps | vs_x500 | vs_x400 | x400_vs_x500 | quad_L7_L8 | two_precision_110_140 | count_A | count_B | planted(48) | evaluator wall s |
|---|---|---|---|---|---|---|---|---|---|
| 2^-6 | 76.33 | 76.33 | 120.41 | 80.09 | 37.83 | 76 | 37 | 46.65 (FAIL by name: True) | 2052.72 |
| 2^-7 | 76.39 | 76.39 | 120.15 | 79.54 | 37.58 | 76 | 37 | 47.01 (FAIL by name: True) | 2019.98 |
| 2^-8 | 76.54 | 76.54 | 119.87 | 79.45 | 37.60 | 76 | 37 | 46.68 (FAIL by name: True) | 2014.10 |
| 2^-9 | 76.75 | 76.75 | 119.58 | 79.56 | 37.76 | 76 | 37 | 47.06 (FAIL by name: True) | 2057.79 |
| 2^-10 | 77.01 | 77.01 | 119.29 | 79.76 | 37.99 | 77 | 37 | 46.96 (FAIL by name: True) | 2065.47 |
| 2^-11 | 77.28 | 77.28 | 118.99 | 80.01 | 38.26 | 77 | 38 | 47.26 (FAIL by name: True) | 1966.32 |
| 2^-12 | 77.57 | 77.57 | 118.70 | 80.29 | 38.54 | 77 | 38 | 47.39 (FAIL by name: True) | 2004.40 |
| 2^-13 | 77.87 | 77.87 | 118.41 | 80.58 | 38.83 | 77 | 38 | 46.86 (FAIL by name: True) | 2111.62 |

- The members vs_x500 and vs_x400 are equal at every eps to the printed precision (the two grids agree with each other far beyond the evaluator's agreement with either); `I_disp_40d` and `O500_40d` in the compare receipt print the two values to 40 significant digits.
- THE PLANTED-DIGIT CONTROL (`per_eps[k].planted_control`): one significant digit of the dps-140 value changed at position 48; the agreement with the x500 grid falls to 46.65-47.39 digits, under the bar at every eps (`fails_by_name` true, rc 1 in the receipt).
- THE TWINS (`grids/`): two AMFlow black-box runs of the same 3 target integrals at the same 15 eps samples (eps = 2^-k), the target sample `result[0]` = the integral with indices [1, 1, 1, 2, 1, 2, 1, 1]; auxiliary-mass order 400 (`options.x_order` 400, `extra_x_order` 440, `working_pre` 340) and 500 (`options.x_order` 500, `extra_x_order` 540, `working_pre` 340).  Their leg receipts (`grids/*_RECEIPT_*.json`): x400 wall 10:34.19, peak RSS 2910832 kB, reductions stored 28 / cache hits 0 (the cold run: every reduction computed and stored); x500 wall 1:25.90, peak RSS 1124976 kB, stored 0 / hits 28 (the hot run: every reduction a cache hit); both under memory.max 5836177408 B, cpu.max `800000 100000`, nice 10; each names its output by sha256 (`out.sha256` = the vendored grid's sha256) and its configuration by sha256 with the numeric values (s, t, msq) = (-3/2, -1/4, 1); `DONE.rule`: `rc 0 AND out parses AND not STANDIN AND 3 results AND the config's 15 eps samples each AND no empty Arb ball AND result[0] the target [1,1,1,2,1,2,1,1]`.
- THE EVALUATOR RECEIPTS (`evals/EVAL_EVAL_P2_2em<k>_dps140_L7_<stamp>_MAIN.json`, one per eps): `I_disp` at 140 digits (`I_disp_sig_digits` 150), the one-loop box w'-DE the run consumed (`box_de`: 8 masters, 46 IBP samples, 0 mismatches; the DE's sha256 `3a33954bfe1a5800be03239bb475bdb6482db6c35c64ca6b03440d594eb971d5` -- the same system as the served `vendor_row33_points/points/P2/box1eq_de.json` (sha256 `8112bb4ce60df93dbc28c2c0cf4315604295df662cf7dd6db76aaff99b5ca650`): equal on every key the served copy carries (the 8 masters, the connection, the 46 samples), the record's copy carrying one extra directory-path field (`samples_dir`) that the served copy does not), `diag.quad_agree` = the L7/L8 pair, `diag.nK` / `tK` / `nR` / `tR` the transport counts and times, `wall_s` 1966.32-2111.62 s each (one core; the recorded setting, not re-run for this shelf).
- THE STATEMENT OBJECT OF RECORD (`statements/STATEMENT_ROW33PP_v2_20260910T045317Z.json`; v1 beside): `state` `PASS at >= 50 d at all eight eps at P2 (convention A; the controls FAIL by name)`; the two printed clauses it carries (`paste.cell.TO`, `paste.sentence_dps110.TO`):
    - `$\ge37$ @ the same $8$ $\eps$-points at four further $(s,t)$ points ($\ge56$\,d at two of them at a second precision: the oracle's truncation floor; $\ge76$ at $P_2$ against higher-order twins)`
    - `a level set by the quadrature at dps~$110$ and not by the oracle's truncation (at $P_2$ the same value at dps~$140$ against a pair of independent \amflow{} grids at auxiliary-mass orders $400$ and $500$ agrees to $76$--$77$ digits, measured $76.33$--$77.87$, the pair itself agreeing to $118.41$ digits at worst)`
  its `NOT_ESTABLISHED` list, in its order (items 1, 2 and 4 as re-cut by field for the public register, 2026-09-11): (1) the three other further points P1, P3, P4 at the higher-order twins: not run (the later runs' wall-time estimates are taken from this pilot's walls); (2) a dps-170 twin of the evaluator (the L1 two-precision member of the P3 / P4 points record, beyond dps 140): not run; under that rule the count at P2 stays >= 37 (the dps-110-vs-dps-140 pair); (3) the oracle dump's integrity scan by oracle_integrity.py (the black_box samples form is not read by the tool); (4) the served lbl3vp-points-evaluate.py carries the x_order-200 grids; the higher-order (x_order 400 / 500) grids are not part of that script; they are a later addition and sit on this shelf as data.
- THE TIMING LINE (`statements/PRICE_LINE_ROW33PP_20260910T034641Z.json`): the measured classes the runs' wall times were estimated from (`price_line_hostB_twins`, `price_line_hostA_evals`: the x_order-200 grid legs of record 317.87 s cold / 16.2887 s hot at P2; the sixteen dps-140 evaluator legs at P3 / P4, walls 1448.47-1599.19 s); the two-hour estimate rule it states (`estimate_gate`).
- THE DPS-110 MEMBER: the compare receipt's `two_precision_110_140` was computed by the comparator from the record's full dps-110 strings (its `record_dps110_dir`, a directory that is not on any shelf); the only served source of the dps-110 value is `../vendor_row33_points/points/P2/record.json` (sha256 `b9960411bca7c40d0235d91223b93e2c5bd2e9b1803c475a7b3fff74ad9a4f4d`), whose `fixed_eps.per_eps[k].I_disp_40d` carries it to 40 significant digits -- the driver below recomputes the member from those strings and reports it AS REPRODUCED TO THE PRECISION A 40-DIGIT STRING ALLOWS, with the difference from the receipt's figure; the served record's own `agreement_d` (vs the x_order-200 grid, 37.58-38.83) is the served page's 37-38 figure.

THE OBJECTS (22 files: the 17 rows of the manifest of record `ROW33_P2_BUNDLE_MANIFEST_20260910T045828Z.txt` -- itself vendored as a re-cut text object, the eighteenth object -- with README.md, VENDOR_MANIFEST.sha256, the reproduction driver `reproduce_row33_p2.py` and the served copy of its run receipt `reproduce_row33_p2-run_20260910T051906Z.json`; original file names, the two statement objects served under the names STATEMENT_ROW33PP_v1_<stamp>.json / STATEMENT_ROW33PP_v2_<stamp>.json (the vendored_from lines below name their emitted sha256).  LAYOUT: `grids/` the two grids + their leg receipts; `evals/` the eight evaluator receipts; `compare/` the compare receipt + the comparator `compare_ROW33PP.py` as a TEXT object (two docstring lines re-cut; it is not run from this shelf: its inputs are the record's paths); `statements/` the two statement objects + the timing line; the manifest of record at the root (its rows by basename, three comment lines re-cut to prose).  Every JSON object re-cut BY FIELD: the compute host's name and the statement objects' lint-capture fields read 'withheld', a path field reads the file name + sha256 (or (directory) / (absent at the cure)), key names unchanged, box labels and the numerical terms as emitted; the two grids byte-identical to the emitted objects (no field moved).

REPRODUCE FROM THIS DIRECTORY (python3 with mpmath; seconds; a fresh --out directory outside the bundle each run -- the driver refuses an existing one or one inside the bundle; the served `../vendor_row33_points/points/P2/record.json` must sit beside this directory as served: the driver pins it by sha256 and refuses by name on a mismatch, exit 2):

    python3 reproduce_row33_p2.py --out <fresh directory>

What it does: every vendored object is read only after its sha256 equals its VENDOR_MANIFEST row (a mismatch refuses, exit 3); the eight receipts are located through the vendored_from header by the emitted sha256 the compare receipt names; the two grids are asserted equal to the compare receipt's `grids.*.sha256` and the leg receipts' `out.sha256`.  It recomputes per eps vs_x500, vs_x400, x400_vs_x500 (the comparator's `agree()` at mp.dps 400 on the full strings, copied verbatim and marked), reads quad_L7_L8 from the receipt's `diag.quad_agree`, forms count_A = int(min(...)) (never rounded) and the planted-digit control at position 48 (`plant()` copied verbatim), and compares every one of those leaves with the compare receipt as the float printed there (the tolerance: equality of the doubles); it prints the cell figure (int(min) over the eight counts), the range and the twins' worst pair, recomputes the dps-110 member from the served record's 40-digit strings (reported with the difference; not an exit condition), and writes REPRODUCE_ROW33_P2.json under --out.  Exit 0 iff every convention-A leaf equals the receipt's and the planted control FAILS by name at every eps.

THE RUN OF RECORD (`reproduce_row33_p2-run_20260910T051906Z.json`, the driver's receipt from a copy of this bundle on 2026-09-10, served beside; its own copy of the run's figures): `reproduced` True; leaves differing 0; the cell figure 76, the range 76-77; vs the x500 grid 76.33-77.87; the L7/L8 pairs 79.45-80.58; the twins' pair 118.41 at worst, 120.41 at best; the planted controls FAIL by name at every eps: True; the dps-110 member from the served 40-digit strings 37.58-38.81 (differences from the receipt's full-string figures -2.81e-02..1.71e-02); mpmath 1.3.0, python 3.12.3.  (The VENDOR_MANIFEST.sha256 named inside that receipt is the emission BEFORE the receipt itself joined the group; the rows the run consumed are listed in it by sha256 and equal the rows served now.)

SERVE ORDER.  This directory is an overlay built ON the bundle's first-manifest staging (its MANIFEST.sha256 `e6484861ecb359df`, 33 rows; its CHANGES.md `08c86e5ea334993c`), which serves FIRST; the bundle's MANIFEST.sha256 re-emitted with this group's rows and CHANGES.md with this entry follow it.  Applied on a bundle without that manifest, the re-emitted MANIFEST.sha256 still lists every served file (the generator walks the directory) -- the CHANGES.md would then carry the first manifest's entry without the emission it names.

Pins: no served script pins this group (lbl3vp-points-evaluate.py pins vendor_row33_points/ in-code, lbl3vp-eps0-evaluate.py pins vendor_row33/, lbl3vp-fastcache.json pins the reference-point evaluator and its two data files; none reads anything here); the bundle MANIFEST.sha256 lists every file of this group; VENDOR_MANIFEST.sha256 here lists the same files (the served bytes) with the originals' sha256 in its header; the driver pins the served points/P2/record.json by sha256 `b9960411bca7c40d0235d91223b93e2c5bd2e9b1803c475a7b3fff74ad9a4f4d`.

THE VENDORED OBJECTS AS EMITTED (one line per object: the shelf path, the sha256 the manifest of record carries for it (the manifest of record itself: its own sha256 as emitted), what moved at the re-cut, the served sha256):
- vendored_from: `grids/LBL3VP_P2_x400.json` <- sha256 `ca2d9da80000c4bd68e0f88600f0e111c24115debecada5ca8366e453ef12911` (no field moved (byte-identical to the emitted object); now sha256 `ca2d9da80000c4bd68e0f88600f0e111c24115debecada5ca8366e453ef12911`)
- vendored_from: `grids/LBL3VP_P2_x400_RECEIPT_20260910T042548Z.json` <- sha256 `c60217ef962184e7f4fea3bb1551316031fd077df3f0f61f4c39404b9f14a99c` (3 field(s) withheld, 9 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `c5e9f95186426ab4b1d94819a2dc3759a2e68342a6d60f1a6c143a554c8c2d89`)
- vendored_from: `grids/LBL3VP_P2_x500.json` <- sha256 `2814cc748cef813bc08b50740133d5c9f5102269ef8a7d906c26703bb929e6d2` (no field moved (byte-identical to the emitted object); now sha256 `2814cc748cef813bc08b50740133d5c9f5102269ef8a7d906c26703bb929e6d2`)
- vendored_from: `grids/LBL3VP_P2_x500_RECEIPT_20260910T042751Z.json` <- sha256 `df9ba1234489f94a82af29a5de3aacc233218211e631883960c70b3890a40da5` (3 field(s) withheld, 9 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `d12036922b7b7b2f767313034d78d242bc230fe4ae2c46e221a20d2d2d175aa9`)
- vendored_from: `evals/EVAL_EVAL_P2_2em10_dps140_L7_20260910T041414Z_MAIN.json` <- sha256 `a1649a491c7e5c0b78a752e5a293b4e2cb9fe68faf7aea780fe4af01ae699362` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `2e1aa2e43714ccb5a1c33169137f0ca519d951e3d2f0a2c9fb723a0646b09d50`)
- vendored_from: `evals/EVAL_EVAL_P2_2em11_dps140_L7_20260910T041418Z_MAIN.json` <- sha256 `53f04f0cdce8a32cc8c494fbad6777b14070940d6c9446c946711d1d5c733ec7` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `749924ff115e46a2191be4f97fb938dc207aec46f1654b75d9cfd9eb36ed631b`)
- vendored_from: `evals/EVAL_EVAL_P2_2em12_dps140_L7_20260910T041422Z_MAIN.json` <- sha256 `ba3867d010b1876ef67c61c08d8a070ee71093bf4eae32fa1a9a14a74a72b5e6` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `d9e1d046c5dbbb69c1637e28221522f5c4f45b1c4132c36434ceaa8f1158f3b5`)
- vendored_from: `evals/EVAL_EVAL_P2_2em13_dps140_L7_20260910T041426Z_MAIN.json` <- sha256 `d9b314e7e76399c751edd3a90dd76bc4e1594d63640efc0cd3ed0edeea73acaf` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `91172b6a88e66c0183f60e7e9b8ce1d2e675d5f41003da800a63adde06cc4ee6`)
- vendored_from: `evals/EVAL_EVAL_P2_2em6_dps140_L7_20260910T041358Z_MAIN.json` <- sha256 `16423d7fc9d61a9146e37c34859d20cc2e3fa329c9a55972cbf3866c78bfd40a` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `7c04ed47976a17c5981e7492cc99692760d7bf751c18274f65c9b89b5cf5cb2c`)
- vendored_from: `evals/EVAL_EVAL_P2_2em7_dps140_L7_20260910T041402Z_MAIN.json` <- sha256 `fe8d004d8854d28e0fb0f3c09145ddb6b0cbdc7edd2c984fa13e847180ddfdf6` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `652d95db304a04d1fa790315a17f575aeb7f9599a5695f4e18dd89f2a5eb5913`)
- vendored_from: `evals/EVAL_EVAL_P2_2em8_dps140_L7_20260910T041406Z_MAIN.json` <- sha256 `f59b97a3c84c9d6be2077d52a126ed252a3b04de621d31467daa16450a118f04` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `3d9d0d024f73e0d0e25edb52f5b5b4870536d3bbd17a64b300136f6805a68161`)
- vendored_from: `evals/EVAL_EVAL_P2_2em9_dps140_L7_20260910T041410Z_MAIN.json` <- sha256 `771032cf57f62513fabeefc90c53dcb688bc5d41f56da395d09e538becfa9d95` (1 field(s) withheld, 4 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `1ea8c894f96bebb96e0156b8561c4e73a24b8e0f1568b33798d7078d1bc44ec7`)
- vendored_from: `compare/COMPARE_ROW33PP_P2_20260910T045045Z.json` <- sha256 `721ca91729b00a587fe12840ed8de4c8fccf98803c9c5ff3a6712632184d68bf` (0 field(s) withheld, 12 path field(s) re-cut, 0 phrase(s) re-cut; now sha256 `82ddc18820fb47bb8ed3f0819f74f6c0f74586f2c5bfc8f92103a59105563e72`)
- vendored_from: `compare/compare_ROW33PP.py` <- sha256 `755751d3324ddfa730189d150214acd96b6e99ffc82426daa1650a82d004d4ae` (2 line(s) re-cut, then docstring lines 8 and 10 (the L1 member and the L1 label of the P3 / P4 points record, worded as the three-points shelf's comparator words them) re-cut for the public register (recut_fields_r33p2_r3.py); now sha256 `9b9a8b2b05f48860d8581555f036bca70880294ac4384fdd3c661f4d37d7eb3b`)
- vendored_from: `statements/PRICE_LINE_ROW33PP_20260910T034641Z.json` <- sha256 `32ba9e445764c166398f2cc7633e3c9cd24874bbdfbbbb9e1956af7d0478019c` (1 field(s) withheld, 5 path field(s) re-cut, 2 phrase(s) re-cut; now sha256 `9ea4b15249b319504cd9ea7311f8a140852ac5c599236219a8beb7d8d1250d73`)
- vendored_from: `statements/STATEMENT_ROW33PP_v1_20260910T045109Z.json` <- sha256 `1c89eb1dea7e91e1624136ad2138f23dbe19f2a0a3efdb5f656bddecfe6fe28b` (10 field(s) withheld, 4 path field(s) re-cut, 6 phrase(s) re-cut, then `rule_by_object`, `NOT_ESTABLISHED[0]`, `NOT_ESTABLISHED[1]`, `NOT_ESTABLISHED[3]` and one `sites_swept_not_pasted` value re-cut by field for the public register (recut_fields_r33p2_r3.py); now sha256 `de626b938ec7e30ecc764d8fdc9cf4fe1c42ad7f6cdce9f049c546811d589168`)
- vendored_from: `statements/STATEMENT_ROW33PP_v2_20260910T045317Z.json` <- sha256 `c9274e6bbb1203edee732d0f02e38aa8a172e860603e22cb9ecd57bb8096dc91` (15 field(s) withheld, 8 path field(s) re-cut, 5 phrase(s) re-cut, then `rule_by_object`, `NOT_ESTABLISHED[0]`, `NOT_ESTABLISHED[1]`, `NOT_ESTABLISHED[3]` and one `sites_swept_not_pasted` value re-cut by field for the public register (recut_fields_r33p2_r3.py); now sha256 `20b1d25fe17d08960148406b3d2977c0a01c3b56745ce5030217f06130e178e5`)
- vendored_from: `ROW33_P2_BUNDLE_MANIFEST_20260910T045828Z.txt` <- sha256 `81445ee31b85617c9e32942948f44f672258379559cac07f4cb681c8cb6d25cd` (17 rows to the basename form (the sha256 of each object as emitted kept), 3 comment lines re-cut to prose, 2 rows renamed with the statement objects, then its third comment line re-cut once more for the public register (recut_fields_r33p2_r3.py); now sha256 `323f6e380653a875103cd55de02d3789c4fa6517af34e967265c90117005f645`)
- re-cut after the first serve form, by field, for the public register (recut_fields_r33p2_r3.py, 2026-09-11): `statements/STATEMENT_ROW33PP_v2_20260910T045317Z.json` -- `NOT_ESTABLISHED[0]` (the later runs' wall-time estimates are taken from the pilot's walls), `NOT_ESTABLISHED[3]` (the higher-order grids are not part of the served four-point script; a later addition, on this shelf as data), the `sites_swept_not_pasted` value of the served-page clause (a later addition to the page), and `rule_by_object` with `NOT_ESTABLISHED[1]` (the L1 label and the L1 two-precision member of the P3 / P4 points record, worded as the three-points shelf's statement objects word them) re-worded, no other key or value moved, key order and serialization kept; sha256 `60a4e8ba578652c8` -> `20b1d25fe17d0896`
- re-cut after the first serve form, by field, for the public register (recut_fields_r33p2_r3.py, 2026-09-11): `statements/STATEMENT_ROW33PP_v1_20260910T045109Z.json` -- the same five values re-worded, nothing else moved; sha256 `37e7ae87f29197b9` -> `de626b938ec7e30e`
- re-cut after the first serve form, for the public register (recut_fields_r33p2_r3.py, 2026-09-11): `ROW33_P2_BUNDLE_MANIFEST_20260910T045828Z.txt` -- its third comment line (the compare-section heading) names the timing line; every row byte-identical; sha256 `956fdc3a3f780abc` -> `323f6e380653a875`
- re-cut after the first serve form, for the public register (recut_fields_r33p2_r3.py, 2026-09-11): `compare/compare_ROW33PP.py` -- docstring lines 8 and 10 (the gloss of the dps-110 member and the gloss of convention B) name the L1 member / the L1 label of the P3 / P4 points record as the three-points shelf's comparator does; every other line byte-identical, the syntax tree identical; sha256 `38809b95f2db0afe` -> `9b9a8b2b05f48860`