[agent-kb] Fveritas: the problem factory — a harness that generates, gates, and solves its own research problems - #280

…s 3/3 controls

…ap (QuantumBFS#112), held-out test set calibrated

…unch discipline

…wth metrics

…ved for cluster

…antumBFS#112 battery)

…ief, figures, data

…gistration, and missing-card rule

- build_run.py materializes the gitignored challenge-run folder
  (tracks/agent-kb/results/.../run.json) so /challenge-report can find
  the sawtooth-erosion run; verified build_report.py + render_report.py
  render cleanly and the skill's scan picks up 2 completed figures
- solution README: mentor quickstart table (5 prompts, expected output
  each), requirements.txt, harness-contribution list for PR element 1
- track README: pin the chosen target (issue QuantumBFS#133, anchor QuantumBFS#112) as the
  file itself instructs
- verified end-to-end in a fresh venv: calibration CALIBRATED, demo
  1 survivor/1 deferred/3 dead, sawtooth anchors green

…4 sequential-loop repositioning

- pf/budget.py + pf/round2.py + run_learning_loop.py: round-2 fleet licensed
  by round-1 heuristics; hop budget waste 29% -> 0%, zero deaths, deferred
  card relaunched bigger with next-step recommendation
- tests/test_learning_loop.py anchors (fleet licensing, fingerprint
  non-collision, static fire, round-1 waste anchor 18/63)
- docs/design/problem-generation.md: v3 methodology (badness filter +
  decomposition granularity) + sequential decision loop addendum
- docs/discussion/: Day-4 brainstorm log (repositioning rationale)
- README: 'The loop closes' section + quickstart note; log.md Day 3/4

…QuantumBFS#133 capability checklist

…ors), deferred verdict, QMC routing

- cards/round3/tfim-ratio-sqrt5-001.yaml: frozen gate (|R-sqrt5|/sigma_R, kill 2.0)
- pf/tfim2d.py: chain/square/triangular/honeycomb PBC clusters, even-parity ED
- tests/test_tfim2d.py: dimer exact, classical limit, strong-field bound,
  parity, Jordan-Wigner cross-check (1e-10), Binder limits — TDD, written first
- round 1 died no_solver (registry gate), builder added, round 2 deferred (0.73):
  R = 2.186 +/- 0.068 vs sqrt5; sigma_R 5700x too large -> routed to sign-free QMC
- brief + data + figure + 2 heuristics entries + telemetry both rounds
- INTERFACE.md: tfim_2d registered

… anchors (provenance-tagged)

…clean-room reproduction verified

- ARGUMENT.md: single human-readable doc arguing correctness (Bethe/XDiag
  anchors, blind calibration 5/5+3/3, frozen gates) and usefulness (teeth in
  both directions, learning loop 29%->0% waste, held-out generalization)
- README: link ARGUMENT.md from the top for reviewers
- reproduction verified in a fresh clone + fresh venv: all 5 README prompts
  green, including the optional learning loop

results/ stays out of git per submission rules, so copy the self-contained
report.html (figures embedded as data URIs) + report.json + run.json into
solutions/problem-factory/report/ and link it from the README top line.