PMAT-3795: CRUX chat verb — a two-turn conversation through apr, llama.cpp and ollama (pty-driven), hf and llamafile; judged on the final turn

[noahgift]

CRUX slice 3: the chat verb

Closes #3795. Stacked on #3789 (→ #3767), so this diff is crux-3774-verb-correspondence...HEAD.

keep-open: #3739 stays open; serve, code, thinking ON, consumer rungs and T-1 wiring are later slices. #3794 is the apr chat backend defect this slice found; it's filed, not fixed here.

What it adds

• --verbs run,chat runs a two-turn conversation per model: "What is 2+2?" then "Now multiply that by 3.". It's judged on the final reply (12, which needs the first turn's 4 carried over), and every reply is recorded. chat without run declines, because the positive control lives in run.
• How each engine is driven:
  • apr: apr chat, one user turn per stdin line.
  • llama.cpp and ollama: scripts/lib/crux_pty_chat.py, a stdlib pseudo-terminal driver. It waits for each CLI's prompt marker, sends one turn, captures the reply, and writes the row-contract JSON the judge already reads.
  • hf and llamafile: gen --verb chat (infra-3c's engines, not merged yet).
• Judge: it parses apr's Assistant: transcript. apr's chat cells carry backend_verified: false, because apr chat reports no backend (apr chat --no-gpu still uploads the GGUF to CUDA (VRAM held outside the GPU lock), and apr chat reports no machine-readable backend #3794); not_covered names it. Run-verb token parity isn't claimed for chat.

Why a pty

Measured on lambda:

• fed a pipe, llama-cli loops on empty > prompts until killed;
• ollama run reads the whole pipe as one prompt and answers both turns in one reply.

Neither takes turns from a pipe, so a piped "multi-turn" test of either would have been a single-turn test wearing a label.

Found while building it: #3794

apr chat --no-gpu still uploads the GGUF to CUDA: it holds VRAM outside the GPU lock, while generation does honour --no-gpu. And apr chat reports no machine-readable backend.

Verified

• bash scripts/check_crux_inference_judge.sh: 42 ok, 0 broke. The chat rows are GREEN, RED (apr "9" vs llama.cpp "12"), RED (no Assistant turn), and the turn/backend/parity record. Three chat mutants each break a row.
• A real run on lambda (cpu lane, qwen2.5-coder-1.5b Q4_K_M, --verbs run,chat): PASS, 4 GREEN. In the chat cell apr, llama.cpp@d1d3c3396 and ollama 0.33.2 each gave ["2 + 2 equals 4.", "4 * 3 equals 12."], and no ollama import was left behind.
• bash scripts/guard_tree.sh --no-cargo: 77 checks, 0 failed on 05685a2de. check_no_competing_harnesses.sh: count=0.

🤖 Generated with Claude Code

[noahgift]

Every commit is already in release/0.69.1-batch-2 (git cherry: 0 unfolded). It reaches main through the release batch PR #3835. Closing as folded, not rejected.