feat(loader): source-aware admission — classify and admit before teardown (fixed #674 G2)

[fivetide]

Summary

Fixed version of #674 (G2 source-aware admission). #674 was reverted in a0fca0d6d alongside #668/#673/#676 because it introduced two admission defects that broke production loads:

• CLS-001: classified Qwen3.5/LFM2 by vision_config/model_type config markers, so every Qwen3.5-family text artifact (whose HF config always embeds vision_config) was refused at admission for "missing vision tower".
• self-refusing VMM commit: commit_ordinary re-ran the VMM readiness check after the new model's own VMM tensors existed, counting them as pending teardown and refusing every fresh kv_backend=vmm load.

This PR delivers the same G2 intent — classify once, admit one effective topology, refuse before destructive side effects — without either defect.

What changed

• crates/hipfire-loader/src/admission.rs (new): read-only admit_source — opens the source, classifies arch_id + vision, decides the effective topology (Single/Pipeline/Expert), and refuses no-carrier / ambiguous-carrier / VMM-allowlist / VMM+pp / carrier-pp / EP-arch / lfm2 vision-no-config / DFlash lm-head quant — all before any GPU, VMM, or teardown work. Vision is decided by the tower tensor (classify_vision), preserving the remediation contract 179a20d7f.
• Carrier::admit_topology (new trait method, default + qwen35/cohere2moe/maple/gemma4 overrides) — mirrors each carrier's load-time pp>1 refusal so a refusal fires at admission.
• Load-entry split: load_model_with_gemma4_drafter / load_model_ep_with_kv_mode become thin wrappers that classify-then-load via load_admitted_with_gemma4_drafter / load_model_ep_admitted, which consume the retained SourceAdmission (no re-open, no re-classify).
• Daemon reorder (crates/hipfire-daemon/src/main.rs): daemon topology refusals + admit_source now run before prior-model teardown. On refusal the prior model stays loaded.

How this differs from #674

	#674 (reverted)	this PR
VL classification	config markers (vision_config/model_type)	tower tensor only (tensor_data(...).is_some())
VMM commit	re-checked after own tensors exist (self-refuses)	single pre-allocation ensure_vmm_ready_for_load
refusal ordering	after prior-model teardown	before teardown (prior model preserved)
source handle	re-opened/re-classified at load	retained SourceAdmission consumed once

Verification

• cargo build (workspace): clean. Full workspace test suite: no failures (incl. 5 new admission contract tests).
• Scoped rustfmt on changed files: clean.
• Hardware (gfx1151, isolated HOME):
  • hipfire run qwen3.6:27b → 64/64 layers, KV cache: Q8 vmm (fresh-daemon VMM commit), coherent Paris.
  • hipfire run qwen3.6-35b-a3b.mq4 → coherent linked-list-reverse output.
  • Failed-load probe: a bad-path load (No such file or directory) and a contradictory kv_backend=vmm + pp=2 load both refuse at admission, and a subsequent generate on the prior model returns done in both cases (prior model survives refusal).

Dependencies / nonclaims

• Root of the re-plan: consumes no upstream types (independent of G1's DeviceMesh; G2 admission decides the topology, G5 will bind it to gpus.mesh).
• DimKind::Ep classification is unchanged from master (EP still routes by arch_id 5|6|9|10); no new axis, no manifest planning, no family cutover.

[Kaden-Schutt]

One regression to fix before this lands, found by running the new admission path on hardware against master:

admission.rs:189 refuses kv_backend=vmm for every EP load. Master serves DeepSeek V4 EP with vmm by design.

Master's load_model_ep_with_kv_mode (crates/hipfire-loader/src/lib.rs) only refuses vmm for the arches whose EP loaders don't support it:

10 if kv_backend_kind == KvBackend::Vmm => Err("KV backend 'vmm' requires tp=1")   // MiniMax
5 | 6 if kv_backend_kind == KvBackend::Vmm => Err("KV backend 'vmm' requires tp=1") // Qwen3.5
9 => load_model_ep_ds4(...)                                                         // DS4: vmm allowed

This PR's admit_source:

if kv_backend == KvBackend::Vmm {
    return Err(format!("KV backend '{}' requires tp=1", kv_backend.as_str()));
}

fires for arch 9 too, so a DS4 EP + vmm load that works on 344e17a2e is refused at admission on this branch. The comment above it says "Mirrors load_model_ep_with_kv_mode's arch_id dispatch + VMM refusal" — it doesn't; master's refusal is per-arch. Fix: gate on matches!(arch_id, 5 | 6 | 10), and keep the DS4 arm vmm-capable.

Two smaller things from the same run, non-blocking:

• load_model_ep_admitted documents no-reopen but still passes path into the per-arch EP loaders, so the retained SourceAdmission.source is dropped on the EP path. Either thread the handle through or don't claim no-reopen for EP yet.
• Pre-existing on master, not yours: ornith-1.5:35b-a3b-mq4r loads at tp=4 on 4× gfx1201 but every generate fails with EP arch mismatch (expected dense Qwen TP) — on master and this branch alike. I'll open an issue.

Everything else checked out: with a model resident, four different refusal kinds (pp>1 on a non-Qwen carrier, missing path, vmm on an unsupported carrier, EP on an unsupported arch) leave the prior model loaded and generating byte-identical output to a fresh master load, where master tears it down; pp=1/pp=2 on 0.8B and 27B and EP tp=4 loads are byte-identical to master. The ratchet-raise label is applied for the +21 daemon lines.

[fivetide]

Addressed the blocking regression and the EP-path doc note.

VMM refusal is per-arch now. admit_source gates the EP vmm refusal on matches!(arch_id, 5 | 6 | 10); DeepSeek V4 (9) stays vmm-capable, mirroring load_model_ep_with_kv_mode's per-arch dispatch. Added a unit test locking the contract (ep_vmm_refusal_is_per_arch: 5|6|10 refuse, 9 admitted, non-vmm never refused).

EP no-reopen claim corrected. load_model_ep_admitted and the SourceAdmission.source field doc now state that the per-arch EP loaders re-open path per rank (the retained source is dropped on the EP path; the single/pp route still consumes it). Kept the doc fix rather than threading the handle through the EP loaders, since those are also reached from the non-admission load_model_ep_with_compressor_cache entry.

Smoke-verified on gfx1151 (single device), rebuilt release daemon:

• DS4 --tp 2 --kv-backend vmm now passes admission and proceeds into load_model_ep_ds4 (fails later only at init_tp: device id 1 out of range (count=1) — no second device on this box).
• MiniMax --tp 2 --kv-backend vmm still refuses at admission with KV backend 'vmm' requires tp=1.

The ornith-1.5 EP arch-mismatch note was left untouched as pre-existing.

Requesting another hardware validation run on the multi-GPU setup to confirm the DS4 EP + vmm load that previously refused at admission now serves.

[hipfire-sol]

hw-gate sol prelim

summary: Adds read-only source-aware admission before prior-model teardown: opens and classifies the artifact, resolves its carrier and effective single/pipeline/expert topology, rejects unsupported KV/topology/vision/DFlash combinations, then passes the retained admission into the existing load route. It also raises the daemon source-line policy ceiling from 4155 to 4176; that policy edit does not execute or access external resources.

run_hardware: true
run_hardware_reasons: The Rust changes perform ordinary local model-file reads and existing GPU loader operations; they add no network access, credential reads, process spawning, dependency/toolchain changes, unsafe blocks, or opaque payloads.; The only script change is a plain-text policy threshold increase for daemon line count; it is not executable harness logic.; Hardware is necessary because the new admission rules depend on real artifact arch IDs, tensor inventories, quant types, carrier claims, and effective topology, which unit tests using only classify_vision inputs cannot establish.

routes:

mode	tag	source	why
battery	qwen3.6:27b	bucket	bucket load
battery	ornith-1.5:35b-a3b-mq4r	bucket	bucket load
battery	lfm2.5:1.2b	bucket	bucket load
battery	qwen3.8:27b-mq4-xt	bucket	bucket load

unavailable_routes:

(none)

claim_assessment: The author claims all supported artifacts still load, fresh VMM load no longer self-refuses, EP remains coherent, and admission failures preserve the prior model. The available batteries can prove successful real-artifact classification and generation for dense Qwen, Qwen3.8, LFM2, and EP Ornith. A chain can exercise continued usability, but the specific bad-path and contradictory vmm+pp refusal-preservation claims are proven only if the hardware evidence includes those explicit failed-load transitions followed by successful generation from the prior model.

[hipfire-sol]

hw-gate evidence — 2 lane(s) — verdict fail

lane hiptrx (gfx1201)

hw-gate evidence

field	value
base	8cd15a62b593e950b6fe0082b88bb79d122d1dc2
head	95e0e654e59a3892777a6fc73894006a576c6e43
buckets	load
host gfx	gfx1201
host rocm	7.15.26333-0000000
device	3
runner	hiptrx
daemon_md5	8afc9f1c778c317ffce80679f22bccdf
hipfire_md5	3ea306ec937a936fabeb526e1729cf6d
build_seconds	6.515400648117065
verdict	fail
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

source: bucket sha256_ok: ✅ size_ok: ✅ status: fail reason: battery: harness exit 1

battery — exit 1 seconds 198.1 status fail

reason: harness exit 1

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	132	62	1.8	14.1	False	False	False	True
battery	1		stop	30	0	78	68	451.6	31.1	False	False	False	True
battery	2		stop	31	0	19	16	484.6	31.3	False	False	False	True
battery	3		stop	47	0	31	8	528.6	31.2	False	False	False	True
battery	4		stop	47	0	17	7	532.9	31.2	False	False	False	True

qwen3.6:27b battery turn 0

```python
def longest_substring_without_repeating(s: str) -> str:
    char_index = {}
    start = 0
    max_length = 0
    max_start = 0
    
    for end, char in enumerate(s):
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        if end - start + 1 > max_length:
            max_length = end - start + 1
            max_start = start
            
    return s[max_start:max_start + max_length]
```

qwen3.6:27b battery turn 1

The sky appears blue during the day because sunlight scatters more effectively in shorter blue wavelengths as it interacts with Earth's atmosphere. At sunset, the sun is lower in the sky, meaning its light must travel through a thicker layer of atmosphere to reach your eyes. This extended path scatters away most of the blue light, allowing the longer red and orange wavelengths to dominate the visual field.

qwen3.6:27b battery turn 2

The capital of France is Paris, and the river that runs through it is the Seine.

qwen3.6:27b battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

qwen3.6:27b battery turn 4

17 + 26 = 43

Answer: 43

ornith-1.5:35b-a3b-mq4r

source: bucket sha256_ok: ✅ size_ok: ✅ status: fail reason: battery: harness exit 1

battery — exit 1 seconds 20.1 status fail

reason: harness exit 1

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	134	64	23.1	7.6	False	False	False	True
battery	1		stop	30	0	87	77	105.8	7.4	False	False	False	True
battery	2		stop	31	0	16	13	110.1	10.8	False	False	False	True
battery	3		stop	47	0	22	6	160.2	9.9	False	False	False	True
battery	4		stop	47	0	17	7	158.8	7.4	False	False	False	True

ornith-1.5:35b-a3b-mq4r battery turn 0

```python
def longest_substring(s):
    if not s:
        return ""
    start = 0
    max_start = 0
    max_len = 0
    char_index = {}
    for end, char in enumerate(s):
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        if end - start + 1 > max_len:
            max_len = end - start + 1
            max_start = start
    return s[max_start:max_start + max_len]
```

ornith-1.5:35b-a3b-mq4r battery turn 1

The sky appears blue during the day because sunlight interacts with particles in the atmosphere, and shorter blue wavelengths scatter more easily than longer wavelengths, spreading blue light across the sky. At sunset, the sun sits lower on the horizon, so its light travels through more atmosphere, causing most of the blue light to scatter away before reaching your eyes. This leaves the longer red and orange wavelengths to dominate, painting the sky in warm sunset colors.

ornith-1.5:35b-a3b-mq4r battery turn 2

The capital of France is Paris, and the River Seine runs through it.

ornith-1.5:35b-a3b-mq4r battery turn 3

{"name": "Alice", "age": 34, "city": "Lisbon"}

ornith-1.5:35b-a3b-mq4r battery turn 4

17 + 26 = 43

Answer: 43

lfm2.5:1.2b

source: bucket sha256_ok: ✅ size_ok: ✅ status: fail reason: battery: harness exit 1

battery — exit 1 seconds 20.1 status fail

reason: harness exit 1

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	0	0	120	58	None	145.1	False	False	False	True
battery	1		stop	0	0	68	59	None	145.92	False	False	False	True
battery	2		stop	0	0	15	13	None	145.63	False	False	False	True
battery	3		stop	0	0	26	8	None	146.07	False	False	False	True
battery	4		stop	0	0	25	13	None	146.2	False	False	False	True

lfm2.5:1.2b battery turn 0

```python
def longest_unique_sub(str):
    seen = {}
    start = 0
    max_len = 0
    max_sub = ""
    
    for i, char in enumerate(str):
        if char in seen and start <= seen[char]:
            start = seen[char] + 1
        else:
            if i - start + 1 > max_len:
                max_len = i - start + 1
                max_sub = str[start:i+1]
        seen[char] = i
    return max_sub
```

lfm2.5:1.2b battery turn 1

The sky appears blue during the day because blue light from the sun is scattered in all directions by small air molecules—a phenomenon called Rayleigh scattering. At sunset, the sunlight travels through more atmosphere, so the red wavelengths are diffracted less and reach your eyes directly. This shift in color explains why the sky turns reddish hues during twilight.

lfm2.5:1.2b battery turn 2

The capital of France is Paris, and the Seine River runs through it.

lfm2.5:1.2b battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

lfm2.5:1.2b battery turn 4

Let's calculate the sum: 17 + 26.  
Adding together gives 43.  
Answer: 43

qwen3.8:27b-mq4-xt

source: bucket sha256_ok: ✅ size_ok: ✅ status: fail reason: battery: harness exit 1

battery — exit 1 seconds 20.1 status fail

reason: harness exit 1

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	133	62	26.7	10.6	False	False	False	True
battery	1		stop	30	0	125	107	36.6	27.9	False	False	False	True
battery	2		stop	31	0	19	16	38.1	9.7	False	False	False	True
battery	3		stop	47	0	17	1	55.9	54.2	False	False	False	True
battery	4		stop	47	0	9	3	56.6	5.1	False	False	False	True

qwen3.8:27b-mq4-xt battery turn 0

```python
def longest_substring_without_repeating(s: str) -> str:
    start = 0
    max_length = 0
    max_start = 0
    char_index = {}
    
    for end, char in enumerate(s):
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        if end - start + 1 > max_length:
            max_length = end - start + 1
            max_start = start
    
    return s[max_start:max_start + max_length]
```

qwen3.8:27b-mq4-xt battery turn 1

Sunlight contains all colors of the visible spectrum, but when it enters the atmosphere, the shorter blue wavelengths scatter off gas molecules much more efficiently than longer wavelengths, a process known as Rayleigh scattering. This scattered blue light is dispersed throughout the sky, making it appear blue to our eyes during midday when sunlight travels through a relatively thin layer of atmosphere. At sunset, the sun is low on the horizon, forcing sunlight to pass through a significantly thicker layer of atmosphere, which scatters away almost all the blue light before it reaches us and allows the longer, less-scattered red and orange wavelengths to dominate the view.

qwen3.8:27b-mq4-xt battery turn 2

The capital of France is Paris, and the river that runs through it is the Seine.

qwen3.8:27b-mq4-xt battery turn 3

{"name":"Alice","age":34,"city":"Lisbon"}

qwen3.8:27b-mq4-xt battery turn 4

43

Answer: 43

kernel

not run

lane hipx (gfx1100)

hw-gate evidence

field	value
base	8cd15a62b593e950b6fe0082b88bb79d122d1dc2
head	95e0e654e59a3892777a6fc73894006a576c6e43
buckets	load
host gfx	gfx1100
host rocm	7.15.26333-0000000
device	0
runner	hipx
daemon_md5	8afc9f1c778c317ffce80679f22bccdf
hipfire_md5	3ea306ec937a936fabeb526e1729cf6d
build_seconds	6.459684610366821
verdict	pass
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

source: bucket sha256_ok: ✅ size_ok: ✅ status: pass reason:

battery — exit 0 seconds 32.6 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	133	62	3.2	17.4	False	False	False	True
battery	1		stop	30	0	82	70	441.2	50.3	False	False	False	True
battery	2		stop	31	0	19	16	454.3	50.4	False	False	False	True
battery	3		stop	47	0	31	8	464.3	50.3	False	False	False	True
battery	4		stop	47	0	6	2	469.4	50.5	False	False	False	True

qwen3.6:27b battery turn 0

```python
def longest_substring_without_repeating(s: str) -> str:
    start = 0
    max_len = 0
    max_start = 0
    char_index = {}
    
    for end, char in enumerate(s):
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        if end - start + 1 > max_len:
            max_len = end - start + 1
            max_start = start
            
    return s[max_start:max_start + max_len]
```

qwen3.6:27b battery turn 1

During the day, sunlight passes through a thinner layer of atmosphere, causing shorter blue wavelengths to scatter more widely and dominate the sky's appearance. At sunset, light travels through a thicker portion of the atmosphere, which scatters away most of the blue and green light before it reaches your eyes. This process leaves primarily the longer red and orange wavelengths to pass through, creating the characteristic reddish hues of sunset.

qwen3.6:27b battery turn 2

The capital of France is Paris, and the river that runs through it is the Seine.

qwen3.6:27b battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

qwen3.6:27b battery turn 4

Answer: 43

ornith-1.5:35b-a3b-mq4r

source: bucket sha256_ok: ✅ size_ok: ✅ status: pass reason:

battery — exit 0 seconds 38.2 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	142	66	391.2	15.6	False	False	False	True
battery	1		stop	30	0	91	82	473.1	92.7	False	False	False	True
battery	2		stop	31	0	16	13	498.5	107.9	False	False	False	True
battery	3		stop	47	0	31	8	723.2	113.8	False	False	False	True
battery	4		stop	47	0	17	7	725.8	116.1	False	False	False	True

ornith-1.5:35b-a3b-mq4r battery turn 0

```python
def longest_substring(s):
    if not s:
        return ""
    
    start = 0
    max_start = 0
    max_length = 0
    char_index = {}
    
    for end in range(len(s)):
        char = s[end]
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        if end - start + 1 > max_length:
            max_length = end - start + 1
            max_start = start
    
    return s[max_start:max_start + max_length]
```

ornith-1.5:35b-a3b-mq4r battery turn 1

The sky appears blue during the day because sunlight interacts with gases in the atmosphere through a process called Rayleigh scattering, where shorter blue wavelengths scatter much more readily than longer wavelengths. This scattered blue light then spreads across the sky, reaching our eyes from all directions. At sunset, sunlight travels through more of the atmosphere at a lower angle, allowing the blue light to be scattered away entirely and leaving only the longer red and orange wavelengths to reach your eyes.

ornith-1.5:35b-a3b-mq4r battery turn 2

The capital of France is Paris, and the Seine River runs through it.

ornith-1.5:35b-a3b-mq4r battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

ornith-1.5:35b-a3b-mq4r battery turn 4

17 + 26 = 43

Answer: 43

lfm2.5:1.2b

source: bucket sha256_ok: ✅ size_ok: ✅ status: pass reason:

battery — exit 0 seconds 8.0 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	0	0	103	47	None	229.4	False	False	False	True
battery	1		stop	0	0	73	65	None	229.56	False	False	False	True
battery	2		stop	0	0	15	13	None	230.77	False	False	False	True
battery	3		stop	0	0	26	8	None	230.09	False	False	False	True
battery	4		stop	0	0	20	11	None	229.89	False	False	False	True

lfm2.5:1.2b battery turn 0

```python
def longest_unique_substring(s):
    char_index = {}
    start = 0
    longest = ""

    for i, char in enumerate(s):
        if char in char_index and char_index[char] > start:
            start = char_index[char] + 1
        char_index[char] = i
        if i - start + 1 > len(longest):
            longest = s[start:i+1]

    return longest
```

lfm2.5:1.2b battery turn 1

During the day, the sky appears blue because the blue light from sunlight is scattered more efficiently by the atmosphere due to the shorter wavelength of blue. At sunset, the sun is lower in the sky, so the red wavelengths of light, which travel longer distances, dominate the light reaching our eyes. This shift in wavelength and scattering patterns gives the sun a reddish hue.

lfm2.5:1.2b battery turn 2

The capital of France is Paris, and the Seine River runs through it.

lfm2.5:1.2b battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

lfm2.5:1.2b battery turn 4

Let's calculate the sum: 17 + 26 equals 43.  
Answer: 43

qwen3.8:27b-mq4-xt

source: bucket sha256_ok: ✅ size_ok: ✅ status: pass reason:

battery — exit 0 seconds 23.8 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
battery	0		stop	34	0	143	67	286.2	28.7	False	False	False	True
battery	1		stop	30	0	87	71	422.5	49.0	False	False	False	True
battery	2		stop	31	0	16	13	432.3	49.3	False	False	False	True
battery	3		stop	47	0	31	8	478.7	49.0	False	False	False	True
battery	4		stop	47	0	17	7	479.3	49.2	False	False	False	True

qwen3.8:27b-mq4-xt battery turn 0

```python
def longest_substring_without_repeating(s: str) -> str:
    if not s:
        return ""
    
    char_index = {}
    start = 0
    max_length = 0
    max_start = 0
    
    for end, char in enumerate(s):
        if char in char_index and char_index[char] >= start:
            start = char_index[char] + 1
        char_index[char] = end
        
        if end - start + 1 > max_length:
            max_length = end - start + 1
            max_start = start
    
    return s[max_start:max_start + max_length]
```

qwen3.8:27b-mq4-xt battery turn 1

During the day, shorter-wavelength blue light is scattered more effectively by gas molecules in the atmosphere than other colors, a phenomenon known as Rayleigh scattering, causing the sky to appear blue. At sunset, sunlight travels through a much thicker layer of the atmosphere, scattering away most of the shorter blue and green wavelengths before reaching an observer's eyes. This leaves the longer-wavelength red and orange light to dominate the sky's color.

qwen3.8:27b-mq4-xt battery turn 2

The capital of France is Paris, and the Seine River runs through it.

qwen3.8:27b-mq4-xt battery turn 3

{
  "name": "Alice",
  "age": 34,
  "city": "Lisbon"
}

qwen3.8:27b-mq4-xt battery turn 4

17 + 26 = 43

Answer: 43

kernel

not run

[hipfire-sol]

hw-gate sol verdict

{
  "claim_verdict": "not-exercised",
  "confidence": 0.97,
  "coverage": {
    "gaps": [
      "All four gfx1201 batteries exited 1 and are fixture failures, despite their decoded turns being coherent.",
      "No chain result was produced, so reordered teardown and continued prior-model usability were not exercised.",
      "Neither an invalid-path transition nor the contradictory vmm+pp transition was followed by generation from the prior model.",
      "The Ornith runs used a single configured device and therefore do not establish expert-parallel rank-local reopen or multi-GPU topology behavior.",
      "The daemon line-threshold policy change requires human review."
    ],
    "surfaces_evidenced": [
      "load"
    ],
    "surfaces_touched": [
      "load",
      "policy"
    ]
  },
  "decision": "block",
  "eyeball": [
    "All reported Qwen3.6, Qwen3.8, LFM2.5, and Ornith turns were coherent, answered their prompts, contained every expected substring, and showed no attractor, empty response, runaway generation, or leaked special token.",
    "The gfx1201 Qwen3.6 first turn incurred 18879 ms prefill and all gfx1201 batteries terminated with harness exit 1; inspect the harness logs to determine why otherwise-correct rows were marked failed."
  ],
  "phase": "verdict",
  "rationale": "The gfx1100 batteries show successful coherent loading and generation for all four artifacts, but the gate rules require a block for any fixture failure, and every gfx1201 fixture failed. More importantly, the central transactional claim around crates/hipfire-daemon/src/main.rs:1566 was not tested: there is no failed-load-then-generate transition or chain evidence. The EP behavior introduced at crates/hipfire-loader/src/lib.rs:2930 was also not demonstrated on a multi-device configuration. The evidence therefore neither proves nor disproves the full author claim.",
  "regressions": []
}

Floor: hard=['hw_run_result=failure', "evidence verdict='fail'", 'policy_paths: scripts/leanup-thresholds.txt'] soft=["coverage_gaps: ['All four gfx1201 batteries exited 1 and are fixture failures, despite their decoded turns being coherent.', 'No chain result was produced, so reordered teardown and continued prior-model usability were not exercised.', 'Neither an invalid-path transition nor the contradictory vmm+pp transition was followed by generation from the prior model.', 'The Ornith runs used a single configured device and therefore do not establish expert-parallel rank-local reopen or multi-GPU topology behavior.', 'The daemon line-threshold policy change requires human review.']"] model_decision=block final=block

[hipfire-sol]

hw-gate sol verdict block: Mandatory qwen3.8:27b-mq4-xt load coverage failed with harness exit 1 on both gfx1201 and gfx1100; the gate rules require a block for any fixture failure. The successful batteries support ordinary dense, LFM2, and Ornith MoE compatibility, but the central fresh-VMM and refusal-preserves-prior-model claims were not exercised. The policy ceiling change in scripts/leanup-thresholds.txt also independently requires human review.

[Kaden-Schutt]

Gate run 33921475093 failed both hardware lanes, and it is not this PR's fault — it is a gate defect I have now fixed in #712.

The lanes execute pr/scripts/serve_harness.py, i.e. this branch's copy of the instrument that judges it. This branch is based on master from before #703, so its harness has no ATTRACTOR_MIN_WINDOW guard. It therefore flagged the [reason] turn — which answered Answer: 43, 9 tokens, recall=1/1 — as a token attractor:

[reason]t5  finish=stop  ctx=47  gen=9  (ans 3w)  tau=2.33  recall=1/1 !ATTRACTOR | '43\n\nAnswer: 43'

Every row in both lanes' evidence records attractor:false, empty:false, runaway:false, recall_ok:true with correct content (working Python, correct Rayleigh-scattering answer, Paris/Seine, valid JSON) — the fixture-level exit 1 came purely from the stale heuristic. The other three fixtures (qwen3.6:27b battery+chain, ornith-1.5:35b-a3b-mq4r, lfm2.5:1.2b) passed on both lanes.

#712 makes serve_harness.py and redline_daemon_harness.py resolve against the gate checkout, the same way fixtures, registry, prompts and review.py already do. Once it lands I will re-dispatch this PR's gate; no change is needed here.

Rebasing onto current master would also dodge it, but that hides a defect that would keep hitting every PR older than a harness fix.

[hipfire-fable]

announcement: The qwen3.8:27b-mq4-xt failure that blocked this run is not your code: the PR branch's serve_harness.py predates master's ATTRACTOR_MIN_WINDOW short-answer fix, so the coherent 3-token answer '43 / Answer: 43' was scored as a token attractor and the harness exited 1 on both lanes; the same battery passes on your daemon binary with master's harness (exit 0, 5/5 recall, Q8 vmm). Beyond the mandatory routes I exercised the claims directly against the daemon protocol on both your binary and master: six refused loads each followed by a generate on the prior model (yours keeps serving, master says 'no model loaded' for four of them), fresh-daemon VMM plus VMM-to-VMM/contiguous swaps and a DFlash-draft swap, Qwen3.5 pp=2 on two GPUs, and Qwen3.5-MoE EP tp=4 on four GPUs (loads on both, refuses at generate on both — the pre-existing #683 issue). No regression found; static audit confirms the admit_topology default and vision probes mirror every carrier's own refusals. This is a hold rather than a merge because the run carries a leanup ratchet raise (daemon_lines 4155 -> 4176) that needs the maintainer's explicit sign-off — consider hoisting the admission match into a helper so the daemon stops growing — and because the branch needs a rebase onto current master so the gate runs the fixed harness; after that the mandatory routes should pass on their own.

investigation:

| question | route | result | evidence |

|---|---|---|---|

| Is the qwen3.8:27b-mq4-xt harness exit 1 a product regression or a harness artifact? | Read lanes/hw-gate-evidence-{hiptrx,hipx}/hw-gate-logs/qwen3.8-27b-mq4-xt-battery.out; python-imported pr and base serve_harness and evaluated _response_has_attractor('', '43\n\nAnswer: 43'); then cd base && HIP_VISIBLE_DEVICES=0 HIPFIRE_DAEMON_BIN=$HW_GATE_BIN/daemon HIPFIRE_CLI_BIN=$HW_GATE_BIN/hipfire python3 scripts/serve_harness.py --model $HIPFIRE_MODELS_DIR/qwen3.8-27b.mq4-xt --mode battery --max-tokens 256 --thinking off --thinking-effort none --max-think-tokens 0 --home $HIPFIRE_HOME/qwen3.8-battery-masterharness --out ... --prompts-file benchmarks/prompts/hw-gate/serve-battery.json | Both lane .out files end with [reason]t5 ... recall=1/1 !ATTRACTOR \| '43\n\nAnswer: 43' then serve_harness: ... produced a token attractor — a correct, coherent answer. PR harness returns attractor=True for that string, master harness returns False. Rerun on the PR daemon with master's harness: exit=0, attractor=0, recall 5/5, KV cache: Q8 vmm (qwen3.8 tag policy is kv_backend=vmm, so this is also a fresh-daemon VMM commit), t5 = '17 + 26 = 43\n\nAnswer: 43', decode 24-68 tok/s. | DECISION-NOTES.md |

| Does a refused load now leave the prior model usable (the PR's headline claim), and what does master do? | $HW_GATE_EVIDENCE/daemon_probe.py --scenario scenario_refusal.json driving the daemon's native JSON-lines protocol: load qwen3.5:0.8b, generate, then six refused loads (bad path; qwen3.6 vmm+pp=2; lfm2.5 vmm allowlist; lfm2.5 tp=2 EP-arch; qwen3.6 tp=2+pp=2; lfm2.5 pp=2 default admit_topology) each followed by a generate on the prior model and a reload. Run on $HW_GATE_BIN/daemon (device 1) and $HW_GATE_BASE_BIN/daemon (device 2). | PR: all six refusals fire at admission (class=validation) and every subsequent generate returns 'The capital of France is Paris. It is the largest city in France...'; daemon alive, exit 0. Base: after bad-path, vmm+pp2, vmm-allowlist and pp2 refusals the next generate is no model loaded (prior model torn down); only the tp>1 refusals preserve it (deferred EP unload). Refusal strings are identical modulo master's 'load failed: ... GPU: ...' wrapper; error class changes internal -> validation. | refusal-probe-pr.json, refusal-probe-base.json, refusal-probe-AB-summary.txt, *.daemon.stderr |

| Fresh-daemon VMM commit and VMM<->VMM / VMM<->contiguous swaps under the reordered teardown (the #674 self-refusal class), plus DFlash lm-head admission with a real draft. | $HW_GATE_EVIDENCE/daemon_probe.py --scenario scenario_vmmswap.json on $HW_GATE_BIN/daemon device 3: fresh load qwen3.6-27b.mq4 kv_backend=vmm -> gen; swap qwen3.8-27b.mq4-xt vmm -> gen; bad-path and vmm+pp2 refusals -> gen; swap qwen3.5-0.8b contiguous -> gen; swap qwen3.6 vmm with params.draft=qwen36-27b-dflash-mq4.hfq dflash_mode=on -> gen. Also: cd base && HIP_VISIBLE_DEVICES=4 ... serve_harness.py --model qwen3.6-27b.mq4 --mode battery --kv-backend vmm ... (PR binaries). | Every load succeeded (fresh vmm 13.7s, vmm->vmm swap 11.5s, ->contiguous 0.8s, ->vmm+draft 20.8s; stderr: 'KV cache: Asym3 vmm', 'DFlash draft loaded: .../qwen36-27b-dflash-mq4.hfq'), every generate coherent ('The River Seine flows gracefully through the heart of Paris...'), refusals under vmm preserved the loaded model. Battery with --kv-backend vmm: exit 0, 5/5 recall, KV cache: Q8 vmm, decode ~36 tok/s. | vmmswap-probe-pr.json (+ .daemon.stderr), qwen3.6-27b-battery-vmm-prbin.{out,json,serve.log} |

| Does the one carrier with a real pp>1 path (Qwen3.5 HFQ, kept by the admit_topology override) still load and generate on two GPUs, A/B against master? | $HW_GATE_EVIDENCE/daemon_probe.py --scenario scenario_pp2.json: qwen3.5-0.8b pp=2 -> gen, bad-path refusal -> gen, qwen3.6-27b pp=2 -> gen. PR on HIP_VISIBLE_DEVICES=0,1; base on 2,4. | Both binaries: 0.8b pp=2 and 27b pp=2 load and generate coherently; the 0.8b greedy text is byte-identical PR vs base ('The capital of France is Paris. While the capital of the Kingdom of Italy was Rome...'). Under pp=2 the bad-path refusal preserves the model on PR, base answers no model loaded. (Both binaries leak <\|im_end\|> and report finish=None on the pp path — pre-existing, identical on master.) | pp2-probe-pr.json, pp2-probe-base.json (+ .log, .daemon.stderr) |

| Does the EP (tp>1) admission + load_model_ep_admitted dispatch still reach a real multi-GPU load, A/B against master? | cd base && HIP_VISIBLE_DEVICES=0,1,2,4 ... serve_harness.py --model ornith-1.5-35b-a3b.mq4r --mode battery --tp 4 --devices 0,1,2,4 ... with HIPFIRE_DAEMON_BIN=$HW_GATE_BIN (then $HW_GATE_BASE_BIN). tp=2 was tried first. | tp=2: refused 'EP qwen35 MoE requires tp=4' (same string in base lib.rs:3348 / pr lib.rs:3391). tp=4: '[loader] EP load complete: 4 ranks' on BOTH binaries, then every generate fails 'EP arch mismatch (expected dense Qwen TP)' on BOTH (serve.log:197,201). Identical to master — this is the known #683 'loader admits, engine refuses' defect that #687 fixes on staging; not introduced by this PR. | ornith-1.5-35b-a3b-mq4r-battery-ep-tp4-prbin.{out,serve.log}, ...-ep-tp4-basebin.{out,serve.log}, ...-ep-tp2-prbin.{out,serve.log} |

| Static: does the new fail-closed admission refuse anything master loads (pp>1 defaults per carrier, EP arch/VMM gates, vision tower probes, DFlash lm-head check)? | Two read-only scouts over pr/ and base/ crates/hipfire-loader/{lib.rs,carriers.rs,admission.rs}, hipfire-daemon/src/main.rs; spot-checked base carriers.rs:490-507, 780-788, 2253-2261 and pr carriers.rs:539-556, 1538-1577 myself. | Every master carrier load refuses pp>1 for HFQ and Dir (base carriers.rs:173,782,1072,1197,1402,1514,1707,1804,1920,2255) except Qwen35 HFQ (base carriers.rs:503-506 -> load_qwen35_pp), which the PR override keeps (pr carriers.rs:387-400). EP refusal set identical to base lib.rs:2903-2934. probe_vision tensor names and the LFM2 tower-without-config refusal are the same names/condition the carriers use (pr carriers.rs:541-543, 1545-1577). SourceAdmission.has_vision is never consumed. resolve_carrier uses the same REGISTRY/probe as master (base lib.rs:2429-2440). | DECISION-NOTES.md (static section) |

unproven:

• Qwen3.5-VL and LFM2-VL artifacts (tower-tensor classification path, and the LFM2 tower-without-config admission refusal) — no VL fixture is present under $HIPFIRE_MODELS_DIR; verified only statically against the carriers' own probes. A host with a qwen3.5 VL .hfq and an LFM2-VL .hfq (with and without vision_config metadata) would close this.

• DeepSeek V4 (arch 9) EP + vmm and MiniMax (arch 10) EP admission — no ds4/minimax fixture on this host; the per-arch ep_vmm_refusal is covered only by the PR's unit test and by static comparison with base lib.rs:2903-2934.

• The gfx1100 (hipx) lane was not re-run by me; the gate's own hipx run of qwen3.8 failed for the same harness reason (identical '43\n\nAnswer: 43' !ATTRACTOR line) and all other hipx fixtures passed, but the master-harness rerun above is gfx1201 only.

rationale: The hard floor fired on two components. (1) Evidence failure: qwen3.8:27b-mq4-xt battery exit 1 on both lanes. Both lanes' .out files (lanes/hw-gate-evidence-{hiptrx,hipx}/hw-gate-logs/qwen3.8-27b-mq4-xt-battery.out) show all five turns coherent with recall satisfied, and the exit came from the PR checkout's serve_harness _token_attractor (pr:2037) flagging the 3-whitespace-token answer '43\n\nAnswer: 43' (maxfreq 2/3). Master fixed this exact false positive with ATTRACTOR_MIN_WINDOW=8 (base serve_harness.py:2037-2050, citing gate run 33862054891); the PR branch is based on 8cd15a6 and predates it, and the gate runs the harness from the PR tree. The hw-gate.json rows show attractor:false only because the harness raises before writing --out (pr:2675 vs 2738) and stale JSON was picked up. Rerun of the fixture on the PR daemon (md5 8afc9f1c) with master's harness: exit 0, attractor=0, recall 5/5 ($HW_GATE_EVIDENCE/qwen3.8-27b-mq4-xt-battery-prbin-masterharness.out). So Sol's single regression entry is withdrawn. (2) Policy: scripts/leanup-thresholds.txt raises daemon_lines 4155 -> 4176 for a main.rs that grew 4148 -> 4176 (+28) — the admission block is inlined into the load handler rather than hoisted into a helper; that ratchet raise needs the maintainer's label/judgment, which is what keeps this at hold rather than merge-staging. On the behavior the PR actually claims, I closed every one of Sol's coverage gaps on hardware: refused loads (bad path, vmm+pp>1, vmm allowlist, EP arch, tp+pp, default admit_topology) all fire before teardown and the prior model keeps generating on the PR daemon, while the base daemon answers 'no model loaded' after four of the six ($HW_GATE_EVIDENCE/refusal-probe-AB-summary.txt); fresh-daemon VMM commit and VMM<->VMM/contiguous swaps and a swap with an explicit DFlash draft all load and generate coherently (vmmswap-probe-pr.json; qwen3.6-27b-battery-vmm-prbin.out exit 0); Qwen3.5 pp=2 on two GPUs loads and generates byte-identically to master for 0.8b and coherently for 27b (pp2-probe-{pr,base}.json); Qwen3.5-MoE EP tp=4 on four GPUs loads on both binaries and refuses at generate on both ('EP arch mismatch (expected dense Qwen TP)', the pre-existing #683 defect — not this PR's regression). Static audit: no carrier that master loads at pp>1 is refused by the new admit_topology default (only Qwen35 HFQ has a pp path on master, base carriers.rs:503-506, and the override keeps it); probe_vision uses the carriers' own tensor names (pr carriers.rs:541-543, 1545-1577); has_vision is never consumed. One wire change to be aware of: admission refusals now carry class 'validation' with no 'load failed: ... GPU: ...' wrapper where master returned class 'internal'.

hard floor: ['hw_run_result=failure', "evidence verdict='fail'", 'policy_paths: scripts/leanup-thresholds.txt']

soft floor: ["coverage_gaps: ['No explicit fresh-daemon qwen3.6:27b load with kv_backend=vmm exercised the VMM self-refusal fix.', 'The ordinary chain battery did not issue a refused replacement load followed by generation, so prior-model preservation was not exercised.', 'No contradictory vmm+pp=2 refusal or invalid-path replacement probe was recorded.', 'The daemon line-threshold policy change requires human review.']"]

[hipfire-fable]

The qwen3.8:27b-mq4-xt failure that blocked this run is not your code: the PR branch's serve_harness.py predates master's ATTRACTOR_MIN_WINDOW short-answer fix, so the coherent 3-token answer '43 / Answer: 43' was scored as a token attractor and the harness exited 1 on both lanes; the same battery passes on your daemon binary with master's harness (exit 0, 5/5 recall, Q8 vmm). Beyond the mandatory routes I exercised the claims directly against the daemon protocol on both your binary and master: six refused loads each followed by a generate on the prior model (yours keeps serving, master says 'no model loaded' for four of them), fresh-daemon VMM plus VMM-to-VMM/contiguous swaps and a DFlash-draft swap, Qwen3.5 pp=2 on two GPUs, and Qwen3.5-MoE EP tp=4 on four GPUs (loads on both, refuses at generate on both — the pre-existing #683 issue). No regression found; static audit confirms the admit_topology default and vision probes mirror every carrier's own refusals. This is a hold rather than a merge because the run carries a leanup ratchet raise (daemon_lines 4155 -> 4176) that needs the maintainer's explicit sign-off — consider hoisting the admission match into a helper so the daemon stops growing — and because the branch needs a rebase onto current master so the gate runs the fixed harness; after that the mandatory routes should pass on their own. The hard floor fired on two components. (1) Evidence failure: qwen3.8:27b-mq4-xt battery exit 1 on both lanes. Both lanes' .out files (lanes/hw-gate-evidence-{hiptrx,hipx}/hw-gate-logs/qwen3.8-27b-mq4-xt-battery.out) show all five turns coherent with recall satisfied, and the exit came from the PR checkout's serve_harness _token_attractor (pr:2037) flagging the 3-whitespace-token answer '43\n\nAnswer: 43' (maxfreq 2/3). Master fixed this exact false positive with ATTRACTOR_MIN_WINDOW=8 (base serve_harness.py:2037-2050, citing gate run 33862054891); the PR branch is based on 8cd15a6 and predates it, and the gate runs the harness from the PR tree. The hw-gate.json rows show attractor:false only because the harness raises before writing --out (pr:2675 vs 2738) and stale JSON was picked up. Rerun of the fixture on the PR daemon (md5 8afc9f1c) with master's harness: exit 0, attractor=0, recall 5/5 ($HW_GATE_EVIDENCE/qwen3.8-27b-mq4-xt-battery-prbin-masterharness.out). So Sol's single regression entry is withdrawn. (2) Policy: scripts/leanup-thresholds.txt raises daemon_lines 4155 -> 4176 for a main.rs that grew 4148 -> 4176 (+28) — the admission block is inlined into the load handler rather than hoisted into a helper; that ratchet raise needs the maintainer's label/judgment, which is what keeps this at hold rather than merge-staging. On the behavior the PR actually claims, I closed every one of Sol's coverage gaps on hardware: refused loads (bad path, vmm+pp>1, vmm allowlist, EP arch, tp+pp, default admit_topology) all fire before teardown and the prior model keeps generating on the PR daemon, while the base daemon answers 'no model loaded' after four of the six ($HW_GATE_EVIDENCE/refusal-probe-AB-summary.txt); fresh-daemon VMM commit and VMM<->VMM/contiguous swaps and a swap with an explicit DFlash draft all load and generate coherently (vmmswap-probe-pr.json; qwen3.6-27b-battery-vmm-prbin.out exit 0); Qwen3.5 pp=2 on two GPUs loads and generates byte-identically to master for 0.8b and coherently for 27b (pp2-probe-{pr,base}.json); Qwen3.5-MoE EP tp=4 on four GPUs loads on both binaries and refuses at generate on both ('EP arch mismatch (expected dense Qwen TP)', the pre-existing #683 defect — not this PR's regression). Static audit: no carrier that master loads at pp>1 is refused by the new admit_topology default (only Qwen35 HFQ has a pp path on master, base carriers.rs:503-506, and the override keeps it); probe_vision uses the carriers' own tensor names (pr carriers.rs:541-543, 1545-1577); has_vision is never consumed. One wire change to be aware of: admission refusals now carry class 'validation' with no 'load failed: ... GPU: ...' wrapper where master returned class 'internal'.

[hipfire-sol]

hw-gate sol verdict block: The gfx1100 batteries show successful coherent loading and generation for all four artifacts, but the gate rules require a block for any fixture failure, and every gfx1201 fixture failed. More importantly, the central transactional claim around crates/hipfire-daemon/src/main.rs:1566 was not tested: there is no failed-load-then-generate transition or chain evidence. The EP behavior introduced at crates/hipfire-loader/src/lib.rs:2930 was also not demonstrated on a multi-device configuration. The evidence therefore neither proves nor disproves the full author claim.

[Kaden-Schutt]

Landed on beta as 4d0e0a1b4. This is a fork PR, so it could not be pushed to; it is staged in the main repo as staging/pr-682 — head 95e0e654e merged with beta at a9f4ca860.

That merge had a real semantic conflict, not a generated-map one: #687's ep_admission (now on beta) versus this PR's admit_source, both in crates/hipfire-loader/src/lib.rs. Resolved by keeping your classify-once split (load_model_ep_admitted) and #687's shared ep_unsupported_arch_message(id) for the backstop arm, so the two refusal messages cannot drift apart.

Because that resolution is mine and no gate run covered it, I proved it on hardware (hiptrx gfx1201, daemon md5 99457c71e4e4):

• battery on the canonical qwen3.8-27b.mq4-xt (mq4v2) trunk — 5/5 turns, attractor 0, empty 0, runaway 0, recall satisfied every turn, avg decode 29.5 tok/s
• LFM2 (arch_id=11) --tp 2 refuses with exactly the shared constructor's text at [validation retryable=false rolled_back=false], serve process still alive
• ornith-1.5-35b-a3b (arch_id=6) --tp 2 refuses through fix(admission): refuse unservable model/topology combinations before allocation (#683 family) #687's rule at 32548 MB free / 32624 MB total — i.e. this PR's restructure did not bypass the refusal fix(admission): refuse unservable model/topology combinations before allocation (#683 family) #687 added, which was the actual risk in that conflict
• 48 tests pass in hipfire-loader + hipfire-daemon

Fable's earlier investigation (run 33921475093) closed every coverage gap Sol raised: six refusal classes all fire before teardown with the prior model still generating, while the base daemon answers no model loaded after four of the six. That is the regression this PR removes, and it is worth more than the diff review.

Two floors, both human calls, both made. scripts/leanup-thresholds.txt is a policy path, and the ratchet raise daemon_lines <= 4155 → 4176 is accepted: the logic did land in hipfire_loader::admission, and what grew in the daemon is the call site plus emit_uncorrelated_error boilerplate that has to live at the protocol boundary. If that ceiling is hit again, collapse the repeated emit blocks into a local helper before raising it further.

Recorded as unproven rather than claimed: DS4 (arch 9) / MiniMax (arch 10) EP admission and the Qwen3.5-VL / LFM2-VL tower classification — no such artifact exists on either lane.

Closing as landed via staging; the commits keep your authorship.