…r capture (unblocks SHIP-007 layer-0 bisection)

Triggering observation 2026-04-28: SHIP-007's hypothesis space has been
narrowed by 5 falsified hypotheses (§28, §28.4(a), §31, §32, #1101).
The remaining bug surface is per-element divergence at some specific
stage of layer-0 forward. Aggregate stats — already emitted by
`apr trace --payload` — are insufficient since they can hide per-element
drift behind similar std values.

This contract defines the missing infrastructure: `--save-tensor <stage>`
flag that captures raw F32 tensor values at chosen forward-pass stages,
written as APRT-magic-prefixed binaries that `apr diff --values` can
load directly.

## Stages enumerated (19 total)

embedding, attn_norm, qkv_matmul, qkv_bias, q_post_rope, k_post_rope,
attention, attn_out, post_attn_residual, ffn_norm, ffn_gate, ffn_up,
ffn_silu, ffn_swigl, ffn_out, post_ffn_residual, layer_output,
final_norm, lm_head

## Falsification tests (8)

- 001: --save-tensor flag recognized
- 002: determinism (byte-identical across runs)
- 003: ffn_gate stage produces expected APR-vs-GGUF diff (corroborates #1099)
- 004: APRT header format self-describing
- 005: multi-stage comma-list works
- 006: NaN preservation
- 007: --layer subset compatible
- 008: pv validates

`pv validate` exits 0 (verified).

## Implementation cost

400-600 LOC + 8 tests, multi-day Rust task.

## Linkage to shipping MODEL-1

Once shipped, the SHIP-007 layer-0 bisection completes in one debug
session: run save-tensor in both APR and GGUF formats, apr diff at each
stage, pinpoint the first divergent stage as the actual bug surface.

SHIP-002/005/006/007/008 (5 PARTIALs) all depend on the SHIP-007 fix.
With this tooling, the fix is unblocked.
paiml/qwen2.5-coder-7b-apache-q4k-v1 ships cleanly through both APR
and GGUF backends → MODEL-1 completes.

Status: PROPOSED. Implementation deferred to multi-day Rust task.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…007 layer-0 bisection (ships MODEL-1) (#1102)

* feat(apr-code): add --emit-trace flag (M28 — ccpa-trace.jsonl emission)

Adds `apr code --emit-trace <path>` flag — when set, after the agent
loop completes the runtime writes a 4-record `ccpa-trace.jsonl` file
to `<path>` describing the run.

Format mirrors the schema at
https://github.com/paiml/claude-code-parity-apr/blob/main/contracts/claude-code-parity-apr-v1.yaml
§ trace_schema. The companion-repo `ccpa measure` subcommand (M26)
consumes this file to score apr-code against canonical Claude Code
reference fixtures.

Records emitted:

  1. session_start  — synthetic UUIDv7-shaped session_id derived from
                      the start ts; ts is a timestamp string;
                      cwd_sha256 is a 64-char placeholder (the
                      companion-repo differ normalizes these at compare
                      time).
  2. user_prompt    — turn 0, verbatim text.
  3. assistant_turn — turn 1, single Block::Text carrying the agent's
                      final response text. Tool dispatch / hook /
                      skill records are M29+ enrichment follow-ups.
  4. session_end    — real elapsed_ms + token counts from
                      AgentLoopResult.usage (input_tokens /
                      output_tokens). Real metadata, not stubbed.

Plumbing:
  - commands_enum.rs   — new `emit_trace: Option<PathBuf>` field on
                         the Code variant.
  - dispatch.rs        — threads it into batuta::agent::code::cmd_code.
  - code.rs cmd_code   — accepts the new param + plumbs to
                         run_single_prompt.
  - code.rs run_single_prompt — captures `Instant::now()` at start;
                         after the agent loop returns Ok(r), if the
                         caller passed --emit-trace, calls the new
                         emit_ccpa_trace helper. On write-failure
                         eprintln! a warning but DO NOT fail the
                         agent run.
  - code.rs emit_ccpa_trace — new helper (~85 LOC) that hand-rolls
                         JSONL via serde_json::json! macros (no new
                         dependency on ccpa_trace types).

Tests (4 new in code_tests.rs::emit_trace_tests):
  - emit_writes_4_jsonl_records_with_correct_kinds
  - emit_carries_prompt_and_response_text
  - emit_carries_token_counts_and_elapsed
  - emit_each_record_has_v1_envelope (per-record back-compat
    invariant from the ccpa-trace v2 schema)

Total in agent::code: 50 → 54 tests passing.

Live dogfood:
  $ apr code --emit-trace /tmp/measured.jsonl \
      -p "Show me which CLAUDE.md takes precedence right now"
  $ cat /tmp/measured.jsonl | jq -r '.kind'
    session_start
    user_prompt
    assistant_turn
    session_end
  $ cat /tmp/measured.jsonl | jq -r 'select(.kind=="session_end")'
    {"v":1,"kind":"session_end","turn":1,"stop_reason":"end_turn",
     "elapsed_ms":3295,"tokens_in":44,"tokens_out":1024}

Real elapsed_ms / token counts populated correctly.

Note: the response text from Qwen3-1.7B in the dogfood was gibberish
(<think>-loop pre-existing aprender concern, see PMAT-190). The trace
format is correct; the model behavior is a separate workstream. The
emit-trace flag works regardless of model quality.

Refs:
  - paiml/claude-code-parity-apr#31 (M26 — ccpa measure subcommand
    that consumes this file)
  - paiml/claude-code-parity-apr/contracts/claude-code-parity-apr-v1.yaml
    § trace_schema (the canonical schema)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* feat(apr-code): default to Qwen3-Coder-30B-A3B-Instruct on 24 GB GPUs

Recommends and auto-discovers Qwen3-Coder-30B-A3B-Instruct as the
default model for `apr code` when present. Aligned with the
research write-up at paiml/claude-code-parity-apr / 2026-04-28.

What ships:

  configs/aliases.yaml
    + new short name `qwen3-coder` →
      hf://unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF
    Now `apr pull qwen3-coder` works.

  crates/aprender-registry/src/aliases.rs
    + matching entry in the in-memory AliasRegistry
      (kept in sync with configs/aliases.yaml).

  crates/aprender-orchestrate/src/agent/manifest.rs
    + `~/.cache/pacha/models/` added to model_search_dirs so
      `apr pull`-cached files (content-hashed names) are visible
      to discovery; pair with a friendly symlink in
      `~/.apr/models/` for the preferred-name filter to recognize.
    + new module-level helper `is_preferred_default_model(path)`:
      case-insensitive substring match against a short list of
      recommended-default model names. Order:
        1. qwen3-coder-30b-a3b
        2. qwen3-coder-next
        3. qwen2.5-coder-32b
        4. qwen2.5-coder-14b
    + discover_model + sort_candidates updated to insert
      preferred-name as a sort key BETWEEN validity (still wins
      overall) and newest-mtime. So when a small recently-pulled
      model exists alongside the recommended default, the
      recommended default is selected — fixing the failure mode
      where Qwen3-1.7B (PMAT-190 thinking-loop bug, emits
      gibberish) was being auto-picked over a known-good 30B model.

Tests (5 new in manifest_tests_discovery.rs, 49 → 54 in agent::manifest):
  - preferred_default_recognises_qwen3_coder_30b_a3b
    (any-case, any-quant matching)
  - preferred_default_rejects_small_fallbacks
    (1.7B / 1.5B / 1.1B / 7B all rejected — the 7B Qwen2.5-Coder is
    still useful but we don't anchor it as the recommended-default
    family for 24 GB GPUs)
  - sort_candidates_promotes_preferred_over_newer
    (preferred-name beats newer-but-smaller mtime)
  - sort_candidates_newer_preferred_beats_older_preferred
    (within preferred-names, mtime still tiebreaks)
  - sort_candidates_validity_outranks_preference
    (Jidoka — invalid preferred loses to valid non-preferred)

Live verification (this PR):

  $ apr pull qwen3-coder
    ✓ Downloaded successfully
      Path: /home/noah/.cache/pacha/models/2b88b180a790988f.gguf
      Size: 17.3 GB

  $ ln -s /home/noah/.cache/pacha/models/2b88b180a790988f.gguf \
         /home/noah/.apr/models/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf

  $ apr code -p "ping" --max-turns 1
    Model: Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf (auto-discovered)
    ↑ default-model preference picked correctly.

Known gap (NOT addressed by this PR):

  After auto-discovery picks the model, both apr-serve subprocess
  and embedded inference fail with:

    Error: driver error: inference failed:
           Invalid shape: Tensor 'blk.0.ffn_up.weight' not found

  Qwen3-Coder-30B-A3B is a Mixture-of-Experts model that uses
  per-expert tensor names (`ffn_up_exps`, `ffn_gate_exps`, etc.),
  not the dense `ffn_up.weight` the current realizar GGUF loader
  expects. qwen3moe architecture support is upstream realizar
  work — separate from this PR. The discovery / alias / preferred-
  name selection mechanism is fully ready for when that lands.

  In the interim users hitting the inference error should fall
  back to a dense model — either Qwen2.5-Coder-32B-Instruct
  (also recognized by is_preferred_default_model) or Qwen2.5-Coder-7B.

Refs:
  - Research write-up: paiml/claude-code-parity-apr / chat 2026-04-28
  - Hugging Face: unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF
  - aprender CLAUDE.md § Claude Messages-API proxy spec — same model
    is already declared as the default for `apr serve anthropic`

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* docs(p3): apr-cli-trace-save-tensor-v1 — contract for per-stage tensor capture (unblocks SHIP-007 layer-0 bisection)

Triggering observation 2026-04-28: SHIP-007's hypothesis space has been
narrowed by 5 falsified hypotheses (§28, §28.4(a), §31, §32, #1101).
The remaining bug surface is per-element divergence at some specific
stage of layer-0 forward. Aggregate stats — already emitted by
`apr trace --payload` — are insufficient since they can hide per-element
drift behind similar std values.

This contract defines the missing infrastructure: `--save-tensor <stage>`
flag that captures raw F32 tensor values at chosen forward-pass stages,
written as APRT-magic-prefixed binaries that `apr diff --values` can
load directly.

## Stages enumerated (19 total)

embedding, attn_norm, qkv_matmul, qkv_bias, q_post_rope, k_post_rope,
attention, attn_out, post_attn_residual, ffn_norm, ffn_gate, ffn_up,
ffn_silu, ffn_swigl, ffn_out, post_ffn_residual, layer_output,
final_norm, lm_head

## Falsification tests (8)

- 001: --save-tensor flag recognized
- 002: determinism (byte-identical across runs)
- 003: ffn_gate stage produces expected APR-vs-GGUF diff (corroborates #1099)
- 004: APRT header format self-describing
- 005: multi-stage comma-list works
- 006: NaN preservation
- 007: --layer subset compatible
- 008: pv validates

`pv validate` exits 0 (verified).

## Implementation cost

400-600 LOC + 8 tests, multi-day Rust task.

## Linkage to shipping MODEL-1

Once shipped, the SHIP-007 layer-0 bisection completes in one debug
session: run save-tensor in both APR and GGUF formats, apr diff at each
stage, pinpoint the first divergent stage as the actual bug surface.

SHIP-002/005/006/007/008 (5 PARTIALs) all depend on the SHIP-007 fix.
With this tooling, the fix is unblocked.
paiml/qwen2.5-coder-7b-apache-q4k-v1 ships cleanly through both APR
and GGUF backends → MODEL-1 completes.

Status: PROPOSED. Implementation deferred to multi-day Rust task.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

…SIFY-QW3-MOE-FORWARD-003 (#1127)

## What ships

Adds `crates/apr-cli/tests/qwen3_moe_apr_run_live_falsifier.rs` —
F-QW3-MOE-C22214-001, an integration test that invokes the user-facing
`apr` binary as a subprocess and asserts:

  1. exit 0
  2. stdout contains ≥1 non-whitespace character

against the cached 17.3 GB Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf
with a fresh date-tagged prompt.

This pins the M32c.2.2.2.1.3 dispatch flip (PR #1126,
squash a902eea) in CI / regression-prevention. Without it, a
future regression that re-routed qwen3_moe back to the dense
`run_gguf_generate` path (which produces garbage on MoE weights)
would slip through CI silently — there'd be no signal at the
`apr run` user-facing surface.

## Live evidence (lambda-vector RTX 4090, 2026-04-29)

```
running 1 test
test f_qw3_moe_c22214_001_apr_run_emits_at_least_one_non_whitespace_char ...
F-QW3-MOE-C22214-001: live `apr run` against /home/noah/.cache/pacha/models/2b88b180a790988f.gguf
F-QW3-MOE-C22214-001: elapsed = 130.945370974s
  stdout (first 200B): === APR Run ===

Source: /home/noah/.cache/pacha/models/2b88b180a790988f.gguf

Output:
.

Completed in 130.83s (cached)

  stderr (first 200B): [BOS-FALLBACK] No tokenizer.ggml.bos_token_id in GGUF — using architecture default for 'qwen3moe'
[BOS-FALLBACK] No tokenizer.ggml.bos_token_id in GGUF — using architecture default for 'qwen3moe'

F-QW3-MOE-C22214-001: PASS
ok

test result: ok. 1 passed; 0 failed; 0 ignored
```

Token quality vs llama.cpp Q4_K (cosine on logits) is M32d. This
test asserts ONLY emit/exit-0 — the discharge gate for
FALSIFY-QW3-MOE-FORWARD-003.

## Skip path

CI runners (and any host without the cached GGUF) print:

  F-QW3-MOE-C22214-001: SKIP — no cached Qwen3-Coder GGUF at any of [...]

and return success. Same skip pattern as
`crates/aprender-serve/tests/qwen3_moe_forward_one_token.rs`
(M32c.2.2.2.1.1 in-process forward primitive).

## Contract chain status

  M32a    qwen3-moe-forward-v1 contract scaffold        SHIPPED (#1099)
  M32b    arch-aware FFN load refuses qwen3_moe          SHIPPED (#1100)
  M32c.1+ MoE descriptor load + per-expert byte slicer   SHIPPED
  M32c.2.2.2.1.1 forward_qwen3_moe method                SHIPPED (#1124)
  M32c.2.2.2.1.2 run_qwen3_moe_generate function         SHIPPED (#1125)
  M32c.2.2.2.1.3 dispatch flip + Q4_K_M qtype dispatch   SHIPPED (#1126)
  M32c.2.2.2.1.4 live `apr run` falsifier               THIS PR
  M32d           numerical parity vs llama.cpp           PENDING

After M32d the contract flips DRAFT → ACTIVE_RUNTIME, which
unblocks the companion-repo FALSIFY-CCPA-013 measured tool-dispatch
parity gate.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>