docs(citations): 5 verified backfills + Vashishtha→Jiralerspong bib fix

Verified every arXiv ID against arxiv.org (primary source) before
touching anything. 5 orphan bib entries matched truth and got a single
clean [@bibkey] annotation each. One bib entry had the wrong arXiv
ID attributed to it — surgery below.

Verified backfills (single clean site per key)
----------------------------------------------
  olivas2025proximal    arXiv 2512.12038 → src/proximal/mtp.py:3
  wu2025estimating      arXiv 2509.18484 → src/interference/orthogonal.py:136
  fang2025inward        arXiv 2506.06615 → src/interference/orthogonal.py:252
  kiciman2023causal     arXiv 2305.00050 → docs/guides/llm_dag_family.md:182
  long2023causal        arXiv 2307.02390 → docs/guides/llm_dag_family.md:183

Each paper's author list + title were cross-checked against the arXiv
abstract page and match exactly what paper.bib claims.

Vashishtha ↔ Jiralerspong — wrong arXiv ID pinned on wrong bib
--------------------------------------------------------------
paper.bib had:
  @misc{vashishtha2024causal,
    title={Causal Inference Using LLM-Guided Discovery},
    author={Vashishtha, Aniket and others},
    howpublished={arXiv preprint arXiv:2402.01207},
    ...
  }

But arXiv:2402.01207 is in fact Jiralerspong-Chen-More-Shah-Bengio
"Efficient Causal Graph Discovery Using Large Language Models" (2024).
The real Vashishtha paper ("Causal Inference Using LLM-Guided
Discovery") lives at arXiv:2310.15117, not 2402.01207.

Meanwhile, src + docs already cited Jiralerspong correctly in 3
places (docs/guides/llm_dag_family.md:184 bullet, llm_dag_loop.py:52
docstring, registry.py:4538 reference kwarg), with NO mention of
Vashishtha anywhere — so the vashishtha2024causal entry was a pure
fabrication bound to a wrong arXiv ID.

Fix: delete vashishtha2024causal (never cited, never would be), add
jiralerspong2024efficient with verified metadata for arXiv:2402.01207,
and annotate the 2 safe sites (docs bullet + llm_dag_loop docstring).
registry.py:4538 is a reference= kwarg inside a function call —
UNSAFE_KWARG by the suggester's filter — left untouched.

Separate SURFACING (not fixed here)
-----------------------------------
arXiv API returned 404 for:
  * "Regression-Based Bidirectional Proximal Causal Inference" by
    Shi-Miao-Tchetgen (claimed arXiv 2507.13965 in paper.bib)
  * "Fortified Proximal Causal Inference" by Yang-Schwartz
    (claimed arXiv 2506.13152 in paper.bib)

Those claimed arXiv IDs resolve to DIFFERENT real papers:
  * 2507.13965 = Min-Zhang-Luo, "A regression-based approach for
    bidirectional proximal causal inference in the presence of
    unmeasured confounding"
  * 2506.13152 = Yu-Shi-Tchetgen Tchetgen, "Fortified Proximal
    Causal Inference with Many Invalid Proxies"

Further oddity: src/proximal/{bidirectional,fortified}.py carry BOTH
the fabricated attribution (top-of-docstring) AND the correct Min/Yu
attribution (in result-metadata strings + docs guides). A structural
rewrite to eliminate the fabrications lives in a follow-up commit;
this commit keeps the scope strictly at already-verified metadata.

Verification
------------
  coverage:        65.4% -> 66.7%       (322 -> 328 cited keys)
  orphans:         170   -> 164         (6 keys fully consumed)
  paper.bib size:  492   unchanged      (vashishtha -1 + jiralerspong +1)
  cumulative:      64.0% -> 66.7%       (+2.7pp across 3 batches, 12 bibkeys)

  Gate 0  pytest 81 tests                                  pass
  Gate 1  audit_bib_duplicates --strict   0/0/0            pass
  Gate 2  audit_bib_coverage  --strict-dangling  0 dangling pass
  Gate 3  audit_citations     --strict    316/0/0          pass
  import test (proximal / interference / causal_llm)       pass

Refs verified via arxiv.org primary source (per CLAUDE.md §10).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford

docs/guides/llm_dag_family.md

@@ -179,9 +179,9 @@ edge looks spurious.
 ## References
 
 - Spirtes, Glymour & Scheines (2000). *Causation, Prediction, and Search.*
-- Kıcıman, Ness, Sharma & Tan (2023). "Causal reasoning and large language models." [arXiv:2305.00050](https://arxiv.org/abs/2305.00050).
-- Long, Piché, Zantedeschi, Schuster & Drouin (2023). "Causal discovery with language models as imperfect experts." [arXiv:2307.02390](https://arxiv.org/abs/2307.02390).
-- Jiralerspong, Chen, More, Shah & Bengio (2024). "Efficient Causal Graph Discovery Using Large Language Models." [arXiv:2402.01207](https://arxiv.org/abs/2402.01207).
+- Kıcıman, Ness, Sharma & Tan (2023). "Causal reasoning and large language models." [arXiv:2305.00050](https://arxiv.org/abs/2305.00050). [@kiciman2023causal]
+- Long, Piché, Zantedeschi, Schuster & Drouin (2023). "Causal discovery with language models as imperfect experts." [arXiv:2307.02390](https://arxiv.org/abs/2307.02390). [@long2023causal]
+- Jiralerspong, Chen, More, Shah & Bengio (2024). "Efficient Causal Graph Discovery Using Large Language Models." [arXiv:2402.01207](https://arxiv.org/abs/2402.01207). [@jiralerspong2024efficient]
 
 <!-- AGENT-BLOCK-START: llm_dag_constrained -->
 

paper.bib

@@ -4101,9 +4101,9 @@ @inproceedings{tran2023inferring
   year={2023}
 }
 
-@misc{vashishtha2024causal,
-  title={Causal Inference Using LLM-Guided Discovery},
-  author={Vashishtha, Aniket and others},
+@misc{jiralerspong2024efficient,
+  title={Efficient Causal Graph Discovery Using Large Language Models},
+  author={Jiralerspong, Thomas and Chen, Xiaoyin and More, Yash and Shah, Vedant and Bengio, Yoshua},
   year={2024},
   howpublished={arXiv preprint arXiv:2402.01207},
   doi={10.48550/arXiv.2402.01207}

src/statspai/causal_llm/llm_dag_loop.py

@@ -49,7 +49,7 @@
 
 Jiralerspong, T., Chen, X., More, Y., Shah, V., & Bengio, Y. (2024).
 "Efficient Causal Graph Discovery Using Large Language Models."
-arXiv:2402.01207.
+arXiv:2402.01207. [@jiralerspong2024efficient]
 """
 from __future__ import annotations
 

src/statspai/interference/orthogonal.py

@@ -133,7 +133,7 @@ def network_hte(
 
     References
     ----------
-    Wu & Yuan (arXiv:2509.18484, 2025).
+    Wu & Yuan (arXiv:2509.18484, 2025). [@wu2025estimating]
     """
     cols = [y, treatment, neighbor_exposure, *covariates]
     missing = set(cols) - set(data.columns)
@@ -249,7 +249,7 @@ def inward_outward_spillover(
 
     References
     ----------
-    Fang, Airoldi & Forastiere (arXiv:2506.06615, 2025).
+    Fang, Airoldi & Forastiere (arXiv:2506.06615, 2025). [@fang2025inward]
     """
     cov_list = list(covariates) if covariates else []
     cols = [y, treatment, inward_exposure, outward_exposure, *cov_list]

src/statspai/proximal/mtp.py

@@ -1,6 +1,6 @@
 """
 Proximal Causal Inference for Modified Treatment Policies
-(Olivas-Martinez, Gilbert & Rotnitzky 2025, arXiv 2512.12038).
+(Olivas-Martinez, Gilbert & Rotnitzky 2025, arXiv 2512.12038). [@olivas2025proximal]
 
 Standard PCI estimates the ATE between a treatment level and its
 reference. Modified Treatment Policies (MTP, Haneuse-Rotnitzky 2013)