Port the #534 distinct-id trajectory-eval fix + source-manifest guard to main with a regression test

Goal

Port the #534 distinct-lora_int_id trajectory-eval fix (commit 298877f9c) and its source-manifest guard to main with a regression test, so no future multi-checkpoint vLLM eval can silently serve a stale adapter.

Background

Task #549's audit found the bug class STILL LIVE on main: src/explore_persona_space/experiments/contrastive_neg_geometry_472/eval_trajectory.py:425 constructs LoRARequest(..., lora_int_id=1, ...) inside the checkpoint loop under one LLM(max_loras=1) engine. The fix exists on branches issue-534/issue-555 (commit 298877f9c: distinct ids per checkpoint + assert_source_delta_g_matches_manifest + --eval-only / --fraction-manifest paths) and is NOT an ancestor of main. eval_guard.py is already partially on main, so the port must reconcile partial overlap rather than blind cherry-pick. Source of truth for the fixed driver: issue-534:src/explore_persona_space/experiments/contrastive_neg_geometry_472/eval_trajectory.py (branch-only — do not grep main for it).

Asks

1. Port 298877f9c's driver fix + tests to main (files: the eval_trajectory.py distinct-id fix, scripts/i504_eval_trajectory.py threading, tests/test_i534_source_manifest_guard.py), reconciling the eval_guard.py overlap.
2. Add a regression test that FAILS on the old constant-id construction and PASSES on the fixed one.
3. Bundle the one-line q_test_extended_50.json SHA-pin in scripts/issue532_predictor_stress.py (code hygiene flagged by the same audit).

Acceptance

• git merge-base --is-ancestor <port-commit> main true.
• Ported + new regression tests pass.
• A future multi-checkpoint trajectory eval from main constructs distinct lora_int_id per checkpoint.

Cost

0 GPU-hours; ~1-2 h code work on the VM (no pod).

Provenance

Filed from #549's epm:follow-ups v1 proposal 1 (auto_run: yes, question_relation: substantially-different, execution: filed-only — runs only on manual triage via /issue).