paper: add verified related work (11 refs) + fix Huang->Deng first author

Related-work search (local qmd/gh/LW + Perplexity/Gemini/ChatGPT/Elicit), all
arXiv ids verified HTTP 200, bibtex+abstracts via the bibtex MCP / arXiv scrape:
- gradient-level reward hacking: ackermann2026gradreg (GR), liu2026harve (HARVE)
- deletable-module precedent (pre-dates Cloud): zhou2023securityvectors
- gradient-projection unlearning: shamsian2025orthograd (OrthoGrad), sun2026ogpsa
- C2 generalisation: taylor2025schoolrewardhacks, nishimuragasparian2025rhgeneralize
- weight-space contrastive direction: fierro2025weightarithmetic
- shortcut gradient surgery: cao2026sart; survey: wang2026rewardhackingsurvey
- idea provenance: mallen2025rhinterventions (AF)
Fix: huang2026directional first author is Deng, Wenlong (arXiv 2605.25189);
sync the cold-reader comment to 'Deng et al.'

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

docs/writeup/main.tex

@@ -367,6 +367,7 @@ enough to route one as it forms.
     \bottomrule
   \end{tabular}
 \end{table}
+% TODO hmm a bit hard conceptually. not direction is... random direction? maybe it shoudl be alternative methods idk
 
 \subsection{Long-run convergence}
 
@@ -439,7 +440,7 @@ extraction set, not only the in-distribution mode (Table~\ref{tab:generalisation
 one-liners are in docs/grad\_routing/related\_work.md.}
 \begin{itemize}
   % COMPREHENSION (cold-reader panel 2026-06-03): the keep-vs-remove inversion
-  % takes two reads. State it plainly first: "Huang projects ONTO a clean
+  % takes two reads. State it plainly first: "Deng et al. project ONTO a clean
   % direction; we project a hack direction OUT." Skeptic also flagged: zeroing
   % delta_S_hack at deploy == not-projecting at deploy, so "deletable knob vs
   % only-constrains-training" is thin unless argued; and we never measured