feat(qwen-story): 8-beat E2E narrative + pmat bug-hunt + daily cron

[noahgift]

Summary

A canonical end-to-end "Qwen story" that doubles as a regression gate. Eight beats, one narrative, every core command group — anchored on the Qwen scale ladder (0.5B safetensors → 30B-MoE GGUF) so the story exercises real production scales, not toy fixtures.

What ships

Artifact	Purpose
scripts/qwen-story.sh (336 LOC)	Runnable story; every beat uses OUT=$(cmd); EC=$? (no pipe-then-$?)
contracts/qwen-story-v1.yaml	3 equations + 8 falsifiers (all PASS locally)
README.md § "A Qwen story"	Replaces the flat ## CLI examples block; fixes 2 broken README examples
.github/workflows/qwen-story-daily.yml	04:17 UTC self-hosted GPU cron + workflow_dispatch

The 8 beats

1. Discover (Registry) — apr pull, apr list
2. Trust (QA) — apr qa, apr validate, apr lint
3. Explore (Inspection) — apr inspect, apr tensors, apr tree
4. Adapt (Model ops) — apr export, apr diff, apr convert
5. Use (Inference) — apr run, apr chat, apr code -p
6. Serve (REST) — apr serve run + curl /v1/chat/completions
7. Operate (Profiling) — apr profile, apr gpu, apr serve plan (on 7B Q4K)
8. Scale (MoE) — apr inspect, apr tensors (on 30B-MoE qwen3moe)

Pmat bug-hunt layer

When PMAT_HUNT=1 (default), each beat emits a structured manifest of high-risk untested code in the command-handler modules it just exercised:

-- pmat bug-hunt manifest (run chat code) --
    gap   crates/apr-cli/src/commands/run.rs:resolve_model_alias (impact=42.3)
    churn crates/apr-cli/src/commands/code.rs:dispatch_agent (commits=11)
    fault crates/aprender-serve/src/api/cuda_chat_backend.rs:try_qwen3_moe (unwrap,panic)

The nightly cron diffs this manifest against the prior successful run and opens (or comments on) a tracking issue when growth exceeds 5 lines — so untested branches in command handlers can't accumulate quietly.

README example fixes (caught while building the script)

The story script revealed two more broken examples in README.md ## CLI examples:

• apr profile model.gguf --roofline → there is no --roofline flag; correct usage is just apr profile model.gguf (the description already says "Deep profiling with Roofline analysis")
• apr bench model.gguf --assert-tps 100 → the --assert-tps flag lives on apr qa, not apr bench

The new ## A Qwen story section replaces this block with verified invocations from the runnable script.

Local smoke test

$ bash scripts/qwen-story.sh
... 14 PASS / 2 FAIL / 0 SKIP

Failed beats:
   - B2 apr validate --quality       # closed by #1870 (in flight)
   - B4 apr export                   # closed by #1868 (in flight)

Both failures are expected until the in-flight fixes land. The story is wired correctly — it catches the panic, it catches the validate threshold, it catches each regression class. Once #1868 + #1870 merge, the story will be 16/0/0 on a host with all 4 Qwen models cached.

Falsifier sweep

All 8 falsifiers in qwen-story-v1.yaml PASS:

F-001 PASS  Script exists and is executable
F-002 PASS  Story has 8 beats (8 beatN_<name>() { definitions)
F-003 PASS  Every beat uses OUT=$(cmd); EC=$? pattern
F-004 PASS  PMAT audit wired per beat (8 pmat_hunt calls)
F-005 PASS  Story referenced from README
F-006 PASS  Daily cron exists
F-007 PASS  Bashrs lints clean (0 errors)
F-008 PASS  Beat 7 doesn't run apr qa on 7B (avoids #1864)

Follow-up

A small follow-up PR will add /dogfood Gate 18 that invokes scripts/qwen-story.sh — kept separate to avoid conflict with #1872 which is already adding Gates 13-17 to the dogfood skill.

Test plan

• All 8 contract falsifiers PASS locally
• bashrs lint scripts/qwen-story.sh — 0 errors
• Local smoke run: 14 PASS / 2 expected FAIL (B2 fix(validate): gate --quality threshold on implemented checks (closes #1866) #1870, B4 fix(export): apr export no longer panics on missing num_layers (closes #1865) #1868)
• YAML parses for both contract and workflow files
• Self-hosted runner config per feedback_self_hosted_only.md
• CI: workspace-test, fmt, contracts-lib
• First nightly cron run (will fire ~04:17 UTC)

🤖 Generated with Claude Code