auto-sync 2026-07-04T07:27:40Z workspace

workspace/README.md

@@ -25,9 +25,10 @@ and per-dimension trainmax selector sweeps completed on 2026-07-04; the best
 tanh selector reaches `0.3819` selected success, but its proposal oracle is
 only `0.3819`, so it does not replace the K=16 `env_clip` outcome gate.
 
-All other Markdown files were removed and consolidated into this README. The
-canonical paper is `latex/main.tex` plus `latex/main.pdf`; experiment evidence
-lives in JSON, TeX, logs, configs, and command files under `runs/`.
+All project-authored Markdown files outside vendored/tool caches were removed
+and consolidated into this README. The canonical paper is `latex/main.tex` plus
+`latex/main.pdf`; experiment evidence lives in JSON, TeX, logs, configs, and
+command files under `runs/`.
 
 ## Research Goal
 
@@ -442,8 +443,13 @@ High-value run directories:
 - `runs/chart_object_embeddings_rgb_refs`: object-layout embedding export.
 - `runs/chart_feature_audit*`: feature-source audits.
 - `runs/ctt_residual_full_seed{0,1,2}`: full residual CTT training.
-- `runs/ctt_residual_ot_full_seed{0,1,2}`: planned full residual CTT + OT
-  training family; only promote after the proxy artifacts exist and pass gates.
+- `runs/ctt_residual_ot_full_seed{0,1,2}`: completed full residual CTT + OT
+  ablation from Slurm job `15203117`; it passes the proxy gate but is worse
+  than plain residual CTT on PPTC and mean positive distance.
+- `runs/ctt_residual_no_cycle_full_seed{0,1,2}`: completed no-cycle residual
+  CTT ablation from Slurm job `15204363`; it passes the proxy gate with lower
+  NegativeNear@0.20 but slightly weaker positive support than cycle residual
+  CTT.
 - `runs/ctt_gated_residual_full_seed{0,1,2}`: full gated residual CTT training.
 - `runs/ctt_residual_component_smoke`: real-data CTT training smoke with
   `--no-markdown-report`; logs configured loss weights and per-component loss
@@ -559,12 +565,15 @@ This is a proxy geometry gate, not rollout success. CTT variants pass by
 improving mean distance to target positives while staying within the
 NegativeNear@0.20 safety slack; they do not beat local-atlas on PPTC thresholds.
 The OT variant passes the same proxy gate but does not improve over plain
-residual CTT. The gated residual variant fails the safety gate.
+residual CTT. The no-cycle ablation passes too, but is slightly worse than
+cycle residual CTT on PPTC and mean positive distance. The gated residual
+variant fails the safety gate.
 
 | Method | PPTC@0.20 | PPTC@0.40 | Neg@0.20 | Pos<Neg | MeanPos | Collapse | Gate |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |
 | local-atlas | 0.4058 | 0.6812 | 0.0368 | 0.5998 | 0.7203 | 0.0661 | baseline |
 | CTT residual full | 0.1981 | 0.6087 | 0.0296 | 0.7352 | 0.4509 | 0.0681 | pass |
+| CTT residual no-cycle full | 0.1884 | 0.6039 | 0.0194 | 0.7391 | 0.4577 | 0.0681 | pass |
 | CTT residual+OT full | 0.1787 | 0.5894 | 0.0187 | 0.7284 | 0.4604 | 0.0681 | pass |
 | CTT residual base+context+obs | 0.2464 | 0.6425 | 0.0343 | 0.7717 | 0.4347 | 0.0681 | pass |
 | CTT gated residual full | 0.2319 | 0.6135 | 0.0527 | 0.7248 | 0.4337 | 0.0681 | fail |
@@ -628,15 +637,30 @@ main CTT variant.
 Purpose: test whether the cycle consistency term in `train_ctt.py` improves
 validation proxy support or only adds optimization baggage.
 
-Submitted on 2026-07-04:
+Completed on 2026-07-04:
 
 ```text
 15204363 scripts/slurm/train_ctt_no_cycle_proxy.sbatch array=0-2
-status at first check: PENDING (Priority)
+exit: 0:0 for tasks 0, 1, 2
+elapsed: 00:01:28 to 00:01:41 per task
 ```
 
-Interpretation: no result is claimed until this job completes and
-`runs/ctt_val_proxy_comparison` is regenerated from the run artifacts.
+Full validation result averaged over three seeds:
+
+| Metric | Value |
+| --- | ---: |
+| Rows | 207 |
+| PPTC@0.20 | 0.1884 |
+| PPTC@0.40 | 0.6039 |
+| NegativeNear@0.20 | 0.0194 |
+| Pos<Neg | 0.7391 |
+| Mean positive distance | 0.4577 |
+| Collapse | 0.0681 |
+
+Interpretation: removing cycle reduces NegativeNear@0.20, but it also slightly
+worsens PPTC and mean positive distance versus the cycle residual row. Cycle is
+not a breakthrough ingredient, but the measured ablation supports keeping it in
+the default CTT support model.
 
 ### Best current K=16 train-clean learned selector
 

workspace/latex/main.fdb_latexmk

@@ -1,13 +1,13 @@
 # Fdb version 4
-["bibtex main"] 1783148457 "main.aux" "main.bbl" "main" 1783148458 0
+["bibtex main"] 1783148457 "main.aux" "main.bbl" "main" 1783149882 0
   "./references.bib" 1783037215 3572 027cd2cf27b344368db784320d235a8c ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/bibtex/bst/base/plain.bst" 1292289607 20613 bd3fbfa9f64872b81ac57a0dd2ed855f ""
-  "main.aux" 1783148458 31207 ad1c38770a9cd9af2e7dddcea6ebbd0c "pdflatex"
+  "main.aux" 1783149882 31207 ad1c38770a9cd9af2e7dddcea6ebbd0c "pdflatex"
   (generated)
   "main.bbl"
   "main.blg"
   (rewritten before read)
-["pdflatex"] 1783148458 "main.tex" "main.pdf" "main" 1783148458 0
+["pdflatex"] 1783149881 "main.tex" "main.pdf" "main" 1783149882 0
   "../paper/sections/theory.tex" 1783037322 4023 e8c0485606fdfea0020faae044bf1189 ""
   "../runs/action_bound_audit_rgb_refs/table.tex" 1783102548 379 5a7d42a0f3c58a06c4eeeaa280615fb3 ""
   "../runs/ctt_base_context_obj_val_rollout_comparison/table.tex" 1783103591 2172 16a82c2b9b79cbd8d2afdfca5503a296 ""
@@ -39,10 +39,10 @@
   "../runs/ctt_residual_smoke_proxy/table.tex" 1783037068 479 a3687dfb2376b4f4dbfc05bec09c73c2 ""
   "../runs/ctt_selector_diagnostic_sweep/table.tex" 1783146173 684 a8085f18ac1568d18a9d9130fbe97ada ""
   "../runs/ctt_test_rollout_comparison/table.tex" 1783103576 2171 9cc76279db5d4ffeb30bd3108263eda6 ""
-  "../runs/ctt_val_proxy_comparison/table.tex" 1783148122 1317 7466422514d6c286947a565b06fff56e ""
+  "../runs/ctt_val_proxy_comparison/table.tex" 1783149630 1440 b9468f1d985843f912c6c6b9fec2cf5d ""
   "../runs/ctt_val_rollout_comparison/table.tex" 1783103576 1874 6e2344e80436f3a5295e8d9c128d776d ""
   "../runs/data_accounting/table.tex" 1783034968 472 a35c3885ec2fd2c61f296ca898e5ee0e ""
-  "../runs/paper_ctt_audit/table.tex" 1783148457 371 8d2eac0cee2b96df6766147c3e9a121b ""
+  "../runs/paper_ctt_audit/table.tex" 1783149409 371 8d2eac0cee2b96df6766147c3e9a121b ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1.enc" 1136849721 2900 1537cc8184ad1792082cd229ecc269f4 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/map/fontname/texfonts.map" 1577235249 3524 cb3e574dea2d1052e39280babc910dc8 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/tfm/jknappen/ec/tcrm1000.tfm" 1136768653 1536 e07581a4bb3136ece9eeb4c3ffab8233 ""
@@ -162,10 +162,10 @@
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/web2c/texmf.cnf" 1692823885 39639 818dd8e5a0dce3a72597a7e6bad45d55 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/fonts/map/pdftex/updmap/pdftex.map" 1692821535 5031293 07973d5e761f645996ac32ebf7b8d1f3 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/web2c/pdftex/pdflatex.fmt" 1692821530 1386016 abd9ba6eb53c47bc5de8d672b6ac87d3 ""
-  "main.aux" 1783148458 31207 ad1c38770a9cd9af2e7dddcea6ebbd0c "pdflatex"
+  "main.aux" 1783149882 31207 ad1c38770a9cd9af2e7dddcea6ebbd0c "pdflatex"
   "main.bbl" 1783148458 3120 92555a174bae4ac684748bbe3fae0b8c "bibtex main"
-  "main.out" 1783148458 2647 6c337a545f287573528c7cb52648bc95 "pdflatex"
-  "main.tex" 1783148444 69693 42751dcb5d69b2db58b8ba4f2b58f2f4 ""
+  "main.out" 1783149882 2647 6c337a545f287573528c7cb52648bc95 "pdflatex"
+  "main.tex" 1783149871 69922 67323a11b46b8221d6f77bab1ec394c2 ""
   "tables/car_decomposition.tex" 1783032219 662 bac24c0014c2ba573860ce199b776934 ""
   "tables/main_results.tex" 1783004655 1061 19bfdcc844614aadefc921dae15c700d ""
   "tables/selector_calibration.tex" 1783011198 788 1b54a8eb88519c50862f4acf718798ab ""

workspace/latex/main.log

@@ -1,4 +1,4 @@
-This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  4 JUL 2026 03:00
+This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  4 JUL 2026 03:24
 entering extended mode
  restricted \write18 enabled.
  %&-line parsing enabled.
@@ -536,14 +536,14 @@ sets vary Lipschitz-continuously
 (./tables/source_score_sweep.tex) (./tables/selector_calibration.tex) [6]
 (../runs/ctt_residual_smoke_proxy/table.tex)
 (../runs/ctt_val_proxy_comparison/table.tex) [7]
-Overfull \hbox (21.77454pt too wide) in paragraph at lines 474--481
+Overfull \hbox (21.77454pt too wide) in paragraph at lines 477--484
 \OT1/cmr/m/n/10 base plus gen-er-ated can-di-dates in Man-iSkill, and writes me
 a-sured rows con-sum-able by \OT1/cmtt/m/n/10 scripts/eval[]metrics.py\OT1/cmr/
 m/n/10 .
  []
 
 
-Overfull \hbox (131.08893pt too wide) in paragraph at lines 482--488
+Overfull \hbox (131.08893pt too wide) in paragraph at lines 485--491
 []\OT1/cmr/m/n/10 A one-chart smoke run at K=2 ver-i-fies the mea-sured path an
 d re-store ac-cu-racy (\OT1/cmtt/m/n/10 runs/ctt[]residual[]rollout[]direct[]sm
 oke[]seed0[]v3\OT1/cmr/m/n/10 ),
@@ -552,19 +552,19 @@ oke[]seed0[]v3\OT1/cmr/m/n/10 ),
 (../runs/action_bound_audit_rgb_refs/table.tex) [8]
 (../runs/ctt_base_context_obs_val_noclip_rollout_comparison/table.tex)
 (../runs/ctt_base_context_obs_test_noclip_rollout_comparison/table.tex)
-Underfull \hbox (badness 10000) in paragraph at lines 571--571
+Underfull \hbox (badness 10000) in paragraph at lines 574--574
 []\OT1/cmr/m/n/10 Table 11: []Scaled raw-action val-i-da-tion re-fresh for \OT1
 /cmtt/m/n/10 base[]context[]obs\OT1/cmr/m/n/10 , K=8, with
  []
 
 (../runs/ctt_base_context_obs_val_scaled0215_rollout_comparison/table.tex)
-Underfull \hbox (badness 10000) in paragraph at lines 582--582
+Underfull \hbox (badness 10000) in paragraph at lines 585--585
 []\OT1/cmr/m/n/10 Table 12: []Scaled raw-action held-out test re-fresh for \OT1
 /cmtt/m/n/10 base[]context[]obs\OT1/cmr/m/n/10 , K=8, with
  []
 
 
-Underfull \hbox (badness 1082) in paragraph at lines 582--582
+Underfull \hbox (badness 1082) in paragraph at lines 585--585
 \OT1/cmtt/m/n/10 --execution-action-scale 0.215 \OT1/cmr/m/n/10 and clip-ping d
 is-abled. The global scale im-proves base action-
  []
@@ -581,8 +581,8 @@ is-abled. The global scale im-proves base action-
 (../runs/ctt_base_context_obs_val_rollout_comparison/table.tex)
 (../runs/ctt_base_context_obj_val_rollout_comparison/table.tex)
 (../runs/ctt_test_rollout_comparison/table.tex)
-(../runs/ctt_base_context_obs_test_rollout_comparison/table.tex)
-(../runs/ctt_dominance_val_to_test/table.tex) [10]
+(../runs/ctt_base_context_obs_test_rollout_comparison/table.tex) [10]
+(../runs/ctt_dominance_val_to_test/table.tex)
 (../runs/ctt_learned_dominance_val_to_test/table.tex)
 (../runs/ctt_base_context_obs_learned_dominance_context_val_to_test/table.tex)
 (../runs/ctt_base_context_obs_learned_dominance_context_tangent_success_tanh_va
@@ -593,7 +593,7 @@ est/table.tex)
 le.tex)
 (../runs/ctt_base_context_obs_learned_dominance_source_envclip_train_to_test/ta
 ble.tex) [11]
-Underfull \hbox (badness 3439) in paragraph at lines 978--978
+Underfull \hbox (badness 3439) in paragraph at lines 981--981
 []\OT1/cmr/m/n/10 Table 31: []Held-out test \OT1/cmtt/m/n/10 env[]clip \OT1/cmr
 /m/n/10 mea-sured roll-out at K=16. The ta-ble is gen-er-ated by
  []
@@ -607,107 +607,107 @@ vclip_k16_train_to_test/table.tex)
 (../runs/ctt_base_context_obs_learned_dominance_train_to_test/table.tex)
 (../runs/ctt_base_context_obs_nonlinear_dominance_basic_positive_train_to_test/
 table.tex)
-Overfull \hbox (31.94885pt too wide) in paragraph at lines 1171--1175
+Overfull \hbox (31.94885pt too wide) in paragraph at lines 1174--1178
 []\OT1/cmtt/m/n/10 scripts/export[]cil[]charts.py\OT1/cmr/m/n/10 , \OT1/cmtt/m/
 n/10 scripts/build[]data[]accounting.py\OT1/cmr/m/n/10 , and \OT1/cmtt/m/n/10 s
 cripts/audit[]cil[]charts.py
  []
 
 
-Overfull \hbox (127.11275pt too wide) in paragraph at lines 1177--1181
+Overfull \hbox (127.11275pt too wide) in paragraph at lines 1180--1184
 []\OT1/cmtt/m/n/10 cil/chart[]features.py \OT1/cmr/m/n/10 cen-tral-izes deploym
 ent-visible chart fea-ture con-struc-tion, and \OT1/cmtt/m/n/10 scripts/audit[]
 chart[]feature[]sources.py
  []
 
 
-Overfull \hbox (104.78587pt too wide) in paragraph at lines 1181--1184
+Overfull \hbox (104.78587pt too wide) in paragraph at lines 1184--1187
 []\OT1/cmtt/m/n/10 scripts/slurm/render[]six[]task[]chart[]observations.sbatch 
 \OT1/cmr/m/n/10 and \OT1/cmtt/m/n/10 scripts/slurm/reexport[]rgb[]ref[]cil[]cha
 rts.sbatch
  []
 
 
-Overfull \hbox (23.72661pt too wide) in paragraph at lines 1184--1189
+Overfull \hbox (23.72661pt too wide) in paragraph at lines 1187--1192
 []\OT1/cmtt/m/n/10 scripts/export[]chart[]observation[]embeddings.py \OT1/cmr/m
 /n/10 and \OT1/cmtt/m/n/10 scripts/export[]chart[]object[]embeddings.py
  []
 
 
-Overfull \hbox (129.91359pt too wide) in paragraph at lines 1184--1189
+Overfull \hbox (129.91359pt too wide) in paragraph at lines 1187--1192
 \OT1/cmr/m/n/10 cre-ate de-ter-min-is-tic 32D RGB-stat and 64D RGB object-layou
 t em-bed-dings, and \OT1/cmtt/m/n/10 scripts/slurm/train[]ctt[]feature[]proxy.s
 batch
  []
 
 
-Overfull \hbox (7.97675pt too wide) in paragraph at lines 1189--1198
+Overfull \hbox (7.97675pt too wide) in paragraph at lines 1192--1201
 []\OT1/cmtt/m/n/10 scripts/eval[]ctt[]generated[]rollout.py \OT1/cmr/m/n/10 and
  \OT1/cmtt/m/n/10 scripts/slurm/eval[]ctt[]generated[]rollout.sbatch
  []
 
 
-Overfull \hbox (20.61057pt too wide) in paragraph at lines 1189--1198
+Overfull \hbox (20.61057pt too wide) in paragraph at lines 1192--1201
 \OT1/cmr/m/n/10 train-split cal-i-bra-tion and meta-data load-ing for deploymen
 t-visible chart fea-tures such as \OT1/cmtt/m/n/10 base[]context[]obs\OT1/cmr/m
 /n/10 .
  []
 
 
-Overfull \hbox (49.05014pt too wide) in paragraph at lines 1189--1198
+Overfull \hbox (49.05014pt too wide) in paragraph at lines 1192--1201
 \OT1/cmr/m/n/10 They also record action-bound va-lid-ity la-bels and sup-port r
 aw-action re-play through \OT1/cmtt/m/n/10 --disable-action-clipping
  []
 
 
-Overfull \hbox (2.58342pt too wide) in paragraph at lines 1198--1202
+Overfull \hbox (2.58342pt too wide) in paragraph at lines 1201--1205
 []\OT1/cmtt/m/n/10 scripts/eval[]nonlinear[]dominance[]selector.py \OT1/cmr/m/n
 /10 shares the chart-compatibility feature-loading path
  []
 
-
-Overfull \hbox (11.7557pt too wide) in paragraph at lines 1202--1205
+[13]
+Overfull \hbox (11.7557pt too wide) in paragraph at lines 1205--1208
 []\OT1/cmtt/m/n/10 scripts/audit[]action[]bounds.py \OT1/cmr/m/n/10 pro-duces \
 OT1/cmtt/m/n/10 runs/action[]bound[]audit[]rgb[]refs\OT1/cmr/m/n/10 , which au-
 dits whether
  []
 
-[13]
-Overfull \hbox (6.0625pt too wide) in paragraph at lines 1224--1227
+
+Overfull \hbox (6.0625pt too wide) in paragraph at lines 1227--1230
 []\OT1/cmtt/m/n/10 scripts/eval[]chart[]positive[]memory[]proxy.py \OT1/cmr/m/n
 /10 and \OT1/cmtt/m/n/10 scripts/build[]ctt[]proxy[]comparison.py \OT1/cmr/m/n/
 10 gen-
  []
 
 
-Overfull \hbox (63.81772pt too wide) in paragraph at lines 1227--1232
+Overfull \hbox (63.81772pt too wide) in paragraph at lines 1230--1235
 \OT1/cmr/m/n/10 min-is-tic sum-maries of \OT1/cmtt/m/n/10 delta[]action\OT1/cmr
 /m/n/10 ; \OT1/cmtt/m/n/10 runs/tangent[]reconstruction \OT1/cmr/m/n/10 and \OT
 1/cmtt/m/n/10 runs/tangent[]reconstruction[]rgb[]refs
  []
 
 
-Overfull \hbox (61.283pt too wide) in paragraph at lines 1232--1236
+Overfull \hbox (61.283pt too wide) in paragraph at lines 1235--1239
 []\OT1/cmtt/m/n/10 scripts/audit[]cil[]charts.py \OT1/cmr/m/n/10 writes the lea
 k-age re-ports \OT1/cmtt/m/n/10 runs/leakage[]audit \OT1/cmr/m/n/10 and \OT1/cm
 tt/m/n/10 runs/leakage[]audit[]rgb[]refs\OT1/cmr/m/n/10 ,
  []
 
 
-Overfull \hbox (20.3325pt too wide) in paragraph at lines 1236--1239
+Overfull \hbox (20.3325pt too wide) in paragraph at lines 1239--1242
 []\OT1/cmtt/m/n/10 scripts/train[]utility[]energy.py \OT1/cmr/m/n/10 and \OT1/c
 mtt/m/n/10 scripts/calibrate[]dominance.py \OT1/cmr/m/n/10 im-ple-ment the util
 -ity/scoring
  []
 
 
-Overfull \hbox (3.63927pt too wide) in paragraph at lines 1245--1249
+Overfull \hbox (3.63927pt too wide) in paragraph at lines 1248--1252
 []\OT1/cmtt/m/n/10 scripts/backfill[]paper[]run[]artifacts.py \OT1/cmr/m/n/10 t
 rans-par-ently back-fills non-Markdown run meta-data such
  []
 
 
-Underfull \hbox (badness 10000) in paragraph at lines 1257--1257
+Underfull \hbox (badness 10000) in paragraph at lines 1260--1260
 []\OT1/cmr/m/n/10 Table 38: []Claim-to-artifact au-dit for this draft. The au-d
 it is gen-er-ated by
  []
@@ -724,7 +724,7 @@ Package rerunfilecheck Info: File `main.out' has not changed.
 Here is how much of TeX's memory you used:
  10295 strings out of 480884
  159358 string characters out of 5900692
- 728572 words of memory out of 5000000
+ 730572 words of memory out of 5000000
  27161 multiletter control sequences out of 15000+600000
  414764 words of font info for 71 fonts, out of 8000000 for 9000
  36 hyphenation exceptions out of 8191
@@ -769,7 +769,7 @@ cm/cmtt9.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texm
 f-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb></cvmfs/soft.computecanad
 a.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/cm-super/sfr
 m1000.pfb>
-Output written on main.pdf (34 pages, 439114 bytes).
+Output written on main.pdf (34 pages, 439381 bytes).
 PDF statistics:
  460 PDF objects out of 1000 (max. 8388607)
  393 compressed objects within 4 object streams

workspace/latex/main.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7d3c2093b0e3ea7a5d6c880de64bad3afc8ef4efbb06091933f682210e4725a6
-size 439114
+oid sha256:9740bd18310310f1de94ab1f223714e7724c3c6e99690e012e6487f4dbef1c23
+size 439381

workspace/latex/main.tex

@@ -448,7 +448,10 @@ and keeps NegativeNear@0.20 within the one-point safety slack, but it does not
 beat local-atlas on \pptc{} coverage. The entropic-OT residual ablation also
 passes the proxy gate, but it is worse than plain residual \ctt{} on both
 \pptc{} coverage and mean positive distance; we therefore keep it as an
-ablation rather than the main support result. Gated residual \ctt{} lowers mean positive
+ablation rather than the main support result. Removing cycle consistency lowers
+NegativeNear@0.20 but is also slightly worse than the cycle residual row on
+\pptc{} and mean positive distance, supporting the default cycle term without
+turning it into a separate contribution. Gated residual \ctt{} lowers mean positive
 distance further, but the three-seed NegativeNear@0.20 average exceeds the
 safety slack, so it fails the proxy gate. The honest interpretation is that
 residual \ctt{} is eligible for measured rollout, while the support/safety