feat(demos): end-to-end demonstrations that recover input traps + cold front door (Phase 4)

[Jammy2211]

Phase 4 of the autocti_assistant epic (PyAutoBrain#136). The demonstrations that
prove the assistant works — and the cold-session front door that makes it usable.

Four runnable demonstrations (scripts/demonstrations/)

Each is small enough to run on a laptop in minutes, and each asserts its
result — a clean run is proof the pipeline works, not just that it executes. All
were run for real against the live stack (arctic clocking, real Nautilus fits):

Demo	Proves	Result
demo_1_calibrate_1d	simulate a 1D dataset from known traps, fit, recover the input	density 0.1300 → 0.1323, release 1.2500 → 1.2608 ✅
demo_2_calibrate_imaging_ci	the same in the 2D charge-injection geometry (Clocker2D)	density 0.1300 → 0.1334, release 1.2500 → 1.3123 ✅
demo_3_correct_1d	clocker.remove_cti suppresses the EPER trail	trail power 3.9× lower after correction ✅
demo_4_aggregate_results	reload demo 1 through the aggregator; the model round-trips	recovered density 0.1323 from disk ✅

Two of these caught real issues that I fixed rather than papered over:

• Demo 4 — factor-graph aggregation. ac.agg.CTIAgg and
  max_log_likelihood_instance assume the single-analysis model layout; a
  factor-graph fit stores results per-factor, so the reloaded instance is
  indexed (instance[0].cti...). Demo 4 uses the samples-based path that works.
• Demo 2 — honest recovery. My first over-trimmed version (2 noisy injection
  levels) recovered the density but not the release timescale (1.25 → 2.90).
  That's a genuine under-constraint, not a tolerance issue, so I fixed the data
  (three cleaner injection levels, more live points) rather than loosen the
  assertion.

Cold-session front door

Fixed llms.txt — the file a fresh chat session reads to bootstrap. It described
the assistant as "for doing strong-lens science" and told agents it could
"explain lensing"; now it correctly frames CTI calibration and routes to the
CTI reference. The two canonical cold-start questions are answerable from the repo
alone: install (the arcticpy recipe in ac_setup_environment) and
calibration (ac_fit_cti_model + the calibration-strategy wiki).

Validation

--scope all · --check-citations · --lint-idioms · --check-provenance ·
--check-version all pass. wiki-currency should stay green; boundary skips.

Remaining lensing leak (systemic, flagged not fixed here)

scripts/AGENTS.md still describes lensing SLaM pipelines (/init-slam, MGE
lens light, pixelised source) — it needs a real CTI-pipeline rewrite, which is
beyond a demonstrations PR. It's part of the same reference-side boundary finding
already on #136 (generic files that absorbed lensing content: the audit's
autogalaxy, the PyPI-install shape, _bootstrap_skill.md, and now scripts/AGENTS.md).

Publish gate — legs 1–4

Legs 1–3 stay green; the demonstrations and the cold-front-door fix are the
substance of leg 4. The final act — flipping the repo public — is left for
explicit human approval and is not done in this PR.

Part of #136 (Phase 4, the last phase).