Add CoT baseline solver for H1 comparison

[Gun-Akash]

P2: token-level CoT baseline (H1 comparison)

Implements the CoT baseline for the H1 fairness comparison per the P2 ticket.

What's here

• marc/eval/baselines/cot_baseline.py — CoTSolver (render_prompt → gpt-5 → parse_answer),
  plugs into the existing run_split_eval harness. Reuses perturb_constants and the
  Checker; no separate perturbation script needed.
• marc/eval/baselines/README.md — model, prompt, run config, known limitations.

Results (N=25, perturb_delta=0.1, k=1)

• solve_rate: 1.0 on both splits → generalization_gap 0.0
• perturbation_robustness: 0.0 (in-dist) vs 0.12 (held-out)

Needs input before this is "done":

• @sparsh — results/ is gitignored, so cot_baseline.json isn't in this PR. How do you
  want the results file — direct send, shared drive, or a !results/p2_main/ exception
  like the existing p4_scale one? It's on disk and ready either way.
• @sparsh — confirm N and perturb_delta match what you need for cross-baseline comparability
  (I used the defaults).
• solve_rate saturates at 1.0, so it doesn't separate the splits on this suite. Perturbation
  does show a gap, but flagging that a harder suite (e.g. length_extrapolation, higher n_vars)
  may be needed for solve_rate to be an informative H1 signal.
• @davin — heads up the baseline exercises perturb_constants + Checker on gpt-5 output;
  parse failures come through as NaN (Checker rejects them correctly).

Ticket: akash/p2-baseline

[ImSpxrsh]

@Gun-Akash, The branch is stale dont merge yet. rebase onto main first & CoT uses k=1 (pass@1) but the MARC main eval uses k=4 (pass@4). N and delta match, but MARC getting 4 candidates vs CoT's 1 isn't same H1 comparison either bump CoT to k=4 (4× gpt-5 token usage) or run MARC at k=1 for this table. Your call, but worth settling before the number goes in the paper.