docs(demo): real AI-agent usage scenarios; deprecate stale sales kit - #4

The existing demo/ kit is April-2026 sales-engineer material — DEMO_RUNBOOK.md
and README.md pitch XERJ to enterprise buyers and cite v1.0.0-rc.1, conformance
1305/1329 (now 1360/1363), a since-fixed RC1 ingest race, and "hybrid not yet
producing fused rankings" (run_hybrid does RRF today). It is not a guide to
what this product is FOR: an AI coding agent's tool.

This replaces the framing with measured, reproducible agent scenarios and
flags the old material rather than deleting it.

## Added

* `SCENARIOS.md` — six real an AI coding agent tasks with measured before/after and
  copy-paste reproduction, every number from a run in this repo:
    1. Orient in unfamiliar data — autoindex map, ~890 tokens flat vs growing
    2. Where/why in a large repo — 5.3x fewer tokens at 170k LOC; recall 6/7
       vs grep 3/7 on prose (grep is lexical, misses reworded answers)
    3. Analytics over logs — 6/6 vs grep 4/6 correctness (the gate caught the
       baseline skipping a .gz and mis-parsing a decimal-comma CSV)
    4. Incident drill-down — indexed lookup 0.007s vs grep 7.3s at 1.2 GB
       (1043x), and flat as data grows
    5. Relational without JOIN — the denormalize pattern: 36 bytes vs 15,914
       (442x) for an identical answer
    6. Cross-session memory — 328-byte semantic recall, a capability grep lacks
  Each scenario states plainly when XERJ does NOT help.

* `gate_retrieval.py` + `make_corpus_prose.py` — the retrieval regime, the
  half the analytics gate could not show. On a 92%-prose corpus XERJ scores
  6/7 vs grep 3/7; token parity at 53 KB, widening to 5.3x fewer at 170k LOC.
  Same anti-cheating discipline as gate.py.

* `RESULTS_analytics.txt` / `RESULTS_retrieval.txt` — committed reference runs
  so a reader sees both regimes without running anything.

## The through-line

One law predicts every result and is printed next to every ratio: retrieval
savings scale with the prose fraction; analytics savings scale with record
count but analytics answers are small either way, so there the win is
correctness and latency, not tokens. A single headline number across both
regimes is dishonest, and neither gate will produce one.

## Deprecated (kept, bannered)

DEMO_RUNBOOK.md and README.md now carry a stale-material banner pointing to
SCENARIOS.md and docs/TOKEN_USAGE.md. Refresh their numbers before any external
use.

Not yet audited: demo/playbooks/ (ES-vs-XERJ perf battle plans) and
demo/usecases/ — a follow-up pass. The autoindex usecase evaluation there
remains the most current of the old material.