fix(vram): free weight sidecars on unload; record the Ep mesh axis on EP loads

[Kaden-Schutt]

Summary

Two verified findings from the 2026-09-02 audit (PR #685, README findings 3 and 5), re-landed individually as the audit recommended:

1. Weight sidecar frees. LlamaWeights::free_gpu freed .buf per weight and skipped the PARO rotation and AWQ scale sidecars, leaking one sidecar set per weight per layer on every reload of an AWQ/PARO llama/qwen3 (DeviceBuffer has no Drop). Every WeightTensor now goes through WeightTensor::free_all, mirroring Qwen35Weights::free_gpu / DflashWeights::free_gpu. Qwen2Weights::free_gpu had the same .buf-only pattern and is fixed the same way (no sidecar is allocated for Qwen2 today, but its forward already reads awq_scale).
2. Gpus::init_ep. Every load_model_ep_* called Gpus::init_tp, which records DeviceMesh::rect(Tp, n), so after an EP load mesh.size_of(Ep) == 1. init_ep is layout-identical to init_tp but records rect(Ep, n); the three EP load sites in hipfire-loader use it. Gpus::single keeps DeviceMesh::single() (absent axes read as 1 by design; the existing single-topology test pins that). feat(runtime): add DeviceMesh topology (fixed #673 G1) #681 landed the mesh type with no readers, so this is the fix before the first reader.

Which crate(s) does this touch?

• crates/hipfire-runtime (llama.rs, multi_gpu.rs), crates/hipfire-loader, crates/hipfire-arch-qwen2

Test plan

• cargo test -p hipfire-runtime — 598 pass, incl. new device_mesh_ep_group_and_tp_absent (Ep groups all ranks, size_of(Ep)==n, size_of(Tp)==1, and symmetrically)
• cargo build -p hipfire-runtime -p hipfire-loader -p hipfire-arch-qwen2 clean
• Hardware fault scenario for (1) — load an AWQ or PARO (krot>0) llama/qwen3 artifact, unload/reload N times, watch vram_free_mb in the diag JSON: it should stay flat instead of dropping by one sidecar set per weight per layer per reload. Not run: no AWQ llama-family artifact on hipx/hiptrx at the time; there is no freed-buffer accounting hook to unit-test against.
• hw-gate when ci: hw-gate — two-seat autonomous review rung (Sol decides hardware, Fable decides staging merges) #679 lands

EP examples (ep_minimax.rs, ep_decode_parity.rs, ep_dspark_topology_probe.rs, ep_deepseek4.rs) still call init_tp; dev harnesses with no mesh readers, left for a follow-up.

[hipfire-sol]

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

lane hiptrx (gfx1201)

hw-gate evidence

field	value
base	3bd2914bac09bb6d48b03cf5b0155473ff1bc922
head	a0aa126817b43f5f76accc5b35495d34c07b4e48
buckets	kernel,load,serve
host gfx	gfx1201
host rocm	7.15.26333-0000000
device	3
runner	hiptrx
daemon_md5	e4ac4ad7260938e5c6ba6a2b17efc813
hipfire_md5	29624d0046e39e42824c7e9913863664
build_seconds	40.78373956680298
verdict	pass
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

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

battery — exit 0 seconds 50.5 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	132	62	1.8	14.2	False	False	False	True
battery	1		stop	30	0	82	70	453.2	31.2	False	False	False	True
battery	2		stop	31	0	16	13	486.2	31.4	False	False	False	True
battery	3		stop	47	0	31	8	530.8	31.3	False	False	False	True
battery	4		stop	47	0	17	7	535.1	31.3	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

During the day, sunlight passes through a thinner layer of the atmosphere, allowing shorter blue wavelengths to scatter more efficiently in all directions. At sunset, however, light travels through a much thicker section of the atmosphere, which scatters away most of the blue and green light before it reaches your eyes. Consequently, only the longer red and orange wavelengths remain to penetrate the atmosphere, giving the sky its warm hues.

qwen3.6:27b battery turn 2

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

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: pass reason:

battery — exit 0 seconds 83.9 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	134	62	23.0	7.6	False	False	False	True
battery	1		stop	30	0	79	69	106.6	7.8	False	False	False	True
battery	2		stop	31	0	16	13	110.4	10.6	False	False	False	True
battery	3		stop	47	0	31	8	159.1	11.9	False	False	False	True
battery	4		stop	47	0	17	7	158.9	7.5	False	False	False	True

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

```python
def longest_substring_without_repeating_chars(s: str) -> str:
    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 gases in Earth's atmosphere, and shorter blue wavelengths scatter much more easily than longer wavelengths. At sunset, sunlight must travel through a thicker portion of the atmosphere, which scatters away the blue light before it reaches your eyes. This leaves primarily the longer red and orange wavelengths to reach your eyes directly, creating the characteristic 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: pass reason:

battery — exit 0 seconds 9.0 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	121	58	None	145.43	False	False	False	True
battery	1		stop	0	0	88	77	None	145.21	False	False	False	True
battery	2		stop	0	0	15	13	None	144.23	False	False	False	True
battery	3		stop	0	0	26	8	None	145.25	False	False	False	True
battery	4		stop	0	0	20	11	None	144.93	False	False	False	True

lfm2.5:1.2b battery turn 0

```python
def longest_unique_substring(s):
    seen = {}
    start = 0
    max_len = 0
    max_sub = ""
    
    for i, char in enumerate(s):
        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 = s[start:i+1]
        seen[char] = i
    return max_sub
```

lfm2.5:1.2b battery turn 1

The sky appears blue during the day because sunlight, though white, contains various colors that scatter equally when entering the Earth's atmosphere. This scattering process is known as Rayleigh scattering, which makes the blue light from distant objects more visible during the day. At sunset, the sun is low on the horizon, so the sunlight travels through more atmosphere, scattering the shorter blue wavelengths away from the light beam and leaving mostly red and orange hues visible.

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

chain — exit 0 seconds 20.5 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
chain	0		stop	0	0	238	132	None	143.46	False	False	False	True
chain	1		stop	0	0	150	106	None	137.61	False	False	False	True
chain	2		stop	0	0	82	74	None	133.77	False	False	False	True
chain	3		stop	0	0	84	62	None	130.64	False	False	False	True
chain	4		stop	0	0	53	38	None	128.33	False	False	False	True

lfm2.5:1.2b chain turn 0

```python
def merge_sorted(a, b):
    """
    Merge two pre-sorted lists into a single sorted list without using the built-in sorted() function.
    
    Parameters:
    a (list): First sorted list
    b (list): Second sorted list
    
    Returns:
    list: A new sorted list containing all elements from lists a and b.
    """
    merged = []
    i, j = 0, 0
    # Traverse both lists and add the smaller element to merged
    while i < len(a) and j < len(b):
        if a[i] <= b[j]:
            merged.append(a[i])
            i += 1
        else:
            merged.append(b[j])
            j += 1
    # Append remaining elements from list a
    while i < len(a):
        merged.append(a[i])
        i += 1
    # Append remaining elements from list b
    while j < len(b):
        merged.append(b[j])
        j += 1
    return merged
```

This function merges two sorted lists by comparing elements iteratively while maintaining order.

lfm2.5:1.2b chain turn 1

Here are the steps to solve the problem:

1. Calculate the distance traveled during the first part of the journey (60 mph for 2.5 hours).
2. Calculate the distance traveled during the second part of the journey (40 mph for 1.5 hours).
3. Add both distances to get the total distance.

Let's do the calculations:

- Distance = Speed × Time
- For the first part: Distance = 60 mph × 2.5 hours = 150 miles
- For the second part: Distance = 40 mph × 1.5 hours = 60 miles

Total distance = 150 miles + 60 miles = 210 miles

Final answer: **210 miles**

lfm2.5:1.2b chain turn 2

The seasons on Earth are caused by the tilt of the Earth's axis as it orbits the Sun, resulting in varying amounts of sunlight reaching different parts of the planet at different times of the year. This tilt causes the Northern Hemisphere to be tilted towards or away from the Sun during different seasons, leading to changes in weather and temperature. Consequently, these shifts in sunlight distribution lead to the seasons we experience annually.

lfm2.5:1.2b chain turn 3

One stormy night, the lighthouse keeper stumbled upon a mysterious shell washed ashore near the rocks. Curious and intrigued, he examined the shell, noticing its unusual markings. As he pulled it closer, the keeper heard a faint voice whispering a cryptic message. Suddenly, a bundle of peculiar letters appeared in his hands, leading him on a wild adventure through history and technology.

lfm2.5:1.2b chain turn 4

1. Keep your code modules small and focused.  
2. Use clear and descriptive variable names.  
3. Write comments to explain complex logic.  
4. Maintain consistent formatting and naming conventions.  
5. Test your code thoroughly to catch issues early.

qwen3.8:27b-mq4-xt

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

battery — exit 0 seconds 50.8 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	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

status: pass

report pass: True

lane hipx (gfx1100)

hw-gate evidence

field	value
base	3bd2914bac09bb6d48b03cf5b0155473ff1bc922
head	a0aa126817b43f5f76accc5b35495d34c07b4e48
buckets	kernel,load,serve
host gfx	gfx1100
host rocm	7.15.26333-0000000
device	0
runner	hipx
daemon_md5	e4ac4ad7260938e5c6ba6a2b17efc813
hipfire_md5	29624d0046e39e42824c7e9913863664
build_seconds	45.91004514694214
verdict	pass
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

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

battery — exit 0 seconds 32.4 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	3.3	18.3	False	False	False	True
battery	1		stop	30	0	84	73	441.4	49.8	False	False	False	True
battery	2		stop	31	0	19	16	454.1	49.9	False	False	False	True
battery	3		stop	47	0	31	8	465.1	49.8	False	False	False	True
battery	4		stop	47	0	6	2	464.4	49.9	False	False	False	True

qwen3.6:27b 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]
```

qwen3.6:27b battery turn 1

During the day, sunlight scatters more readily off air molecules when traveling a shorter path through the atmosphere, allowing shorter blue wavelengths to dominate what we see. At sunset, the light travels through a much thicker layer of air, which filters out most of the blue and green light before it reaches your eyes. Consequently, only the longer red and orange wavelengths survive this longer journey, giving the sunset its characteristic warm hue.

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 38.4 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	130	61	458.3	19.9	False	False	False	True
battery	1		stop	30	0	111	97	467.1	29.2	False	False	False	True
battery	2		stop	31	0	16	13	489.7	107.8	False	False	False	True
battery	3		stop	47	0	31	8	720.4	114.2	False	False	False	True
battery	4		stop	47	0	17	7	726.2	116.0	False	False	False	True

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

```python
def longest_substring_no_repeats(s: str) -> str:
    start = max_length = 0
    best_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
            best_start = start

    return s[best_start:best_start + max_length]
```

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

During the day, the sky appears blue because sunlight interacts with atmospheric gases through a process called Rayleigh scattering, in which shorter blue wavelengths scatter much more efficiently than longer wavelengths, filling the sky with scattered blue light from all directions. At sunset, the sun sits low on the horizon, meaning its light travels through a much thicker layer of atmosphere, so nearly all of the easily-scattered blue light is dispersed away before reaching your eyes. This leaves predominantly the longer red and orange wavelengths to reach you directly, creating the characteristic warm hues of a sunset.

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 8.1 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	114	56	None	217.97	False	False	False	True
battery	1		stop	0	0	74	65	None	222.89	False	False	False	True
battery	2		stop	0	0	15	13	None	227.27	False	False	False	True
battery	3		stop	0	0	26	8	None	226.09	False	False	False	True
battery	4		stop	0	0	25	13	None	227.27	False	False	False	True

lfm2.5:1.2b battery turn 0

```python
def longest_unique_substring(s):
    seen = set()
    start = 0
    max_len = 0
    max_sub = ""

    for i in range(len(s)):
        while s[i] in seen:
            start = start ^ 1 if start != i else start
        seen.add(s[i])
        if i - start + 1 > max_len:
            max_len = i - start + 1
            max_sub = s[start:i+1]
    return max_sub
```

lfm2.5:1.2b battery turn 1

The sky appears blue during the day due to Rayleigh scattering, where sunlight interacts with molecules and small particles in the atmosphere, dispersing shorter blue wavelengths more efficiently. At sunset, the angle of the sunlight decreases, so the longer red wavelengths dominate as they travel through more atmosphere. These physical phenomena blend to produce the vibrant colors we see in different times of the day.

lfm2.5:1.2b battery turn 2

The capital of France is Paris, and the River Seine 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

chain — exit 0 seconds 15.3 status pass

mode	idx	genre	finish	ctx	cached	gen	ans_words	prefill_tok_s	decode_tok_s	attractor	empty	runaway	recall_ok
chain	0		stop	0	0	237	127	None	166.78	False	False	False	True
chain	1		stop	0	0	192	121	None	176.31	False	False	False	True
chain	2		stop	0	0	89	74	None	152.66	False	False	False	True
chain	3		stop	0	0	100	78	None	151.29	False	False	False	True
chain	4		stop	0	0	61	44	None	150.99	False	False	False	True

lfm2.5:1.2b chain turn 0

```python
def merge_sorted(a, b):
    """
    Merges two sorted lists into a single sorted list.
    
    Parameters:
    a (list): First sorted list.
    b (list): Second sorted list.
    
    Returns:
    list: A new sorted list containing all elements from both input lists.
    """
    merged = []
    i, j = 0, 0
    
    # Traverse both lists and add the smallest elements
    while i < len(a) and j < len(b):
        if a[i] < b[j]:
            merged.append(a[i])
            i += 1
        else:
            merged.append(b[j])
            j += 1
    
    # Append remaining elements from list a
    while i < len(a):
        merged.append(a[i])
        i += 1
    
    # Append remaining elements from list b
    while j < len(b):
        merged.append(b[j])
        j += 1
    
    return merged
```

This function efficiently merges two sorted lists into a single sorted list without using the built-in `sorted()`.

lfm2.5:1.2b chain turn 1

Certainly! Let's break this down step by step.

**Step 1: Calculate distance for the first part of the trip**  
Train speed = 60 mph  
Time = 2.5 hours  

Distance = Speed × Time  
Distance�� = 60 mph × 2.5 hours = 150 miles

**Step 2: Calculate distance for the second part of the trip**  
Train speed = 40 mph  
Time = 1.5 hours  

Distance = Speed × Time  
Distance�� = 40 mph × 1.5 hours = 60 miles

**Step 3: Add both distances together**  
Total distance = Distance�� + Distance�� = 150 miles + 60 miles = 210 miles

**Final Answer:** The train traveled a total of **210 miles**.

Let me know if you want a Python version of this calculation!

lfm2.5:1.2b chain turn 2

The seasons on Earth are caused by the tilt of the planet's axis relative to its orbit around the Sun, which changes the amount of sunlight different regions receive throughout the year. As the Earth spins and its axial tilt shifts, different parts of the planet are exposed to varying angles of sunlight, leading to seasonal changes. This cyclical pattern results in the distinct seasons—spring, summer, autumn, and winter—over the course of a year.

lfm2.5:1.2b chain turn 3

A reclusive lighthouse keeper lived on a remote island, keeping watch over the waves with a steady beam of light. One rainy morning, while checking the rocks near the shore, he spotted an odd, weathered object half-buried in the sand. Intrigued, he carefully dug it out, revealing a hidden box filled with mysterious letters. As he read the messages, the keeper realized he had uncovered a secret from decades past, changing his understanding of his isolated life forever.

lfm2.5:1.2b chain turn 4

1. Use clear and consistent naming conventions.  
2. Write modular code with well-defined functions.  
3. Document your code with comments explaining complex logic.  
4. Keep each function small and focused on a single task.  
5. Regularly refactor your code to maintain readability and efficiency.

qwen3.8:27b-mq4-xt

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

battery — exit 0 seconds 21.8 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	140	63	310.1	29.0	False	False	False	True
battery	1		stop	30	0	102	87	419.8	48.3	False	False	False	True
battery	2		stop	31	0	19	16	432.6	48.5	False	False	False	True
battery	3		stop	47	0	22	6	478.9	48.2	False	False	False	True
battery	4		stop	47	0	17	7	480.3	48.5	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
    best_start = 0
    best_end = 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 > best_end - best_start:
            best_start = start
            best_end = end
    
    return s[best_start:best_end + 1]
```

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

During the day, sunlight passes through a greater amount of atmosphere, causing shorter-wavelength blue light to scatter in all directions via Rayleigh scattering, which makes the sky appear blue from our perspective. At sunset, the light travels through a significantly thicker layer of the atmosphere, scattering away most of the blue and green light so that only the longer-wavelength red and orange wavelengths remain visible. This differential scattering based on wavelength allows the red light to reach our eyes directly, creating the characteristic hues of a sunset.

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

17 + 26 = 43

Answer: 43

kernel

status: pass

report pass: True

[hipfire-sol]

hw-gate sol verdict

{
  "claim_verdict": "not-exercised",
  "confidence": 0.97,
  "coverage": {
    "gaps": [
      "No repeated AWQ/PARO Llama-family unload/reload VRAM measurement exercised WeightTensor::free_all sidecar reclamation.",
      "No multi-rank EP load exercised Gpus::init_ep or inspected Ep/Tp mesh dimensions.",
      "No Qwen2 fixture exercised its changed unload path.",
      "No DFlash run exercised sidecar resolution, mandatory-draft failure, context exhaustion, emitter rollback, or speculative terminal behavior.",
      "No Gemma4 lowered/MoE artifact exercised the new admission refusal.",
      "CLI pull/rm behavior and registry sidecar policy were not exercised end to end.",
      "The changed registry/models.json and registry/v1.json policy requires human review."
    ],
    "surfaces_evidenced": [
      "ordinary model loading and coherent serving on gfx1100 and gfx1201",
      "Llama-family battery and chain generation",
      "rotated-weight artifact generation",
      "Qwen3.5 kernel capture stability and HIP/PM4 bit-exact parity"
    ],
    "surfaces_touched": [
      "load and unload VRAM lifecycle",
      "multi-GPU EP topology",
      "serve and continuous-batch admission",
      "DFlash loading and terminal state",
      "kernel and dispatch",
      "CLI pull, remove, and benchmark behavior",
      "registry policy"
    ]
  },
  "decision": "needs-human",
  "eyeball": [
    "All qwen3.6:27b, qwen3.8:27b-mq4-xt, ornith-1.5:35b-a3b-mq4r, and lfm2.5:1.2b battery outputs were nonempty, responsive, and free of attractors.",
    "Both lfm2.5:1.2b chain runs remained responsive across related turns; the gfx1100 arithmetic response contains replacement characters in subscript-like labels but still reaches the correct 210-mile answer.",
    "The gfx1100 LFM longest-substring sample contains suspicious application logic despite coherent syntax; this is model-answer quality rather than evidence tied to the changed lifecycle code."
  ],
  "phase": "verdict",
  "rationale": "Hardware evidence is clean for ordinary load/serve and kernel execution: both lanes passed, Qwen3.5 capture sequences were stable, and HIP/PM4 state was bit-exact. It does not prove either principal PR claim: crates/hipfire-runtime/src/llama.rs:686 was not tested with repeated sidecar-bearing unload/reload VRAM accounting, and crates/hipfire-runtime/src/multi_gpu.rs:319 was not exercised with multiple EP ranks. The much broader diff also changes DFlash state handling and registry policy without corresponding routes, so human review is required.",
  "regressions": []
}

Floor: hard=['policy_paths: registry/models.json,registry/v1.json'] soft=["coverage_gaps: ['No repeated AWQ/PARO Llama-family unload/reload VRAM measurement exercised WeightTensor::free_all sidecar reclamation.', 'No multi-rank EP load exercised Gpus::init_ep or inspected Ep/Tp mesh dimensions.', 'No Qwen2 fixture exercised its changed unload path.', 'No DFlash run exercised sidecar resolution, mandatory-draft failure, context exhaustion, emitter rollback, or speculative terminal behavior.', 'No Gemma4 lowered/MoE artifact exercised the new admission refusal.', 'CLI pull/rm behavior and registry sidecar policy were not exercised end to end.', 'The changed registry/models.json and registry/v1.json policy requires human review.']", 'model needs-human'] model_decision=needs-human final=needs-human

[hipfire-sol]

hw-gate sol verdict needs-human: Ordinary model loading and generation passed on both gfx1100 and gfx1201 for qwen3.6:27b, qwen3.8:27b-mq4-xt, ornith-1.5:35b-a3b-mq4r, and lfm2.5:1.2b, with coherent decoded output and no demonstrated regression. However, the two behaviors changed by the PR were not directly established: crates/hipfire-runtime/src/llama.rs:686 lacked repeated sidecar-bearing reload and VRAM evidence, while crates/hipfire-runtime/src/multi_gpu.rs:319 could not run with multiple EP ranks. Multi-GPU topology changes and these coverage gaps require human disposition.

[hipfire-fable]

announcement: Holding on the hard floor, not on your code: the registry/models.json and registry/v1.json edits that tripped the policy rule come in through your 'Merge beta' commit from PR #686, not from any commit you authored, and this head is already on beta from the previous round — the gate re-ran a merged PR. I re-checked the two claims on the 5x gfx1201 host: at this head ornith 4x EP is refused by #687's admission before init_ep ever runs (and master, while it loaded that topology, could not generate on it), so init_ep is now only reachable from DeepSeek V4 / MiniMax loads, which this host cannot run; the reachable 4-rank dense TP path for qwen3.6:27b is byte-identical to master across load/generate/unload/reload. The sidecar free_all change reads correct by ownership (alias guard, per-weight AWQ upload), but there is no AWQ/PARO llama-family or Qwen2 artifact here to measure it, so it stays recorded as unproven rather than proven. Evidence is under fable-evidence/ with a README index.

investigation:

| question | route | result | evidence |

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

| Are the policy-path edits that fired the hard floor (registry/models.json, registry/v1.json) authored by this PR, or inherited from staging? | git diff --stat 3bd2914 3f3d0bc -- registry/ ; git diff --stat 3f3d0bc a0aa126 -- registry/ ; git merge-base --is-ancestor HEAD origin/beta | PR-authored commits (e5a016d, 5c4c7ad, 2cf2cf1, 3f3d0bc) touch only llama.rs, qwen2.rs, multi_gpu.rs, loader/lib.rs and two map.md files; the registry diff against base is EMPTY. Both registry files (+108/-4) enter solely via a0aa126 'Merge beta into fix/vram-leak-relands', i.e. PR #686's dflash-sidecar pairing already on staging. HEAD a0aa126 is already an ancestor of origin/beta (7df822e 'hw-gate: merge PR #688 ... to staging', 21:30 UTC) — this run is a re-gate of a merged head. | README.md (commands and result recorded) |

| Is Gpus::init_ep reachable on this host at this head (the previous round's ornith 4x EP evidence predates the beta merge that brought in #687)? | python3 $HW_GATE_EVIDENCE/ep_probe.py $HW_GATE_BIN/daemon $HIPFIRE_MODELS_DIR/ornith-1.5-35b-a3b.mq4r 4 $HW_GATE_EVIDENCE/ep_admission_ornith_tp4_HEAD.json --devices 0,1,2,3 --generate | Refused in 0.282 s: 'load failed: Qwen3.5-MoE (arch_id=6) has no EP serve path; use TP or single-GPU'. vram_free_mb 32548 before and after; no 'EP load:' stderr line, so the refusal (crates/hipfire-loader/src/lib.rs:3416, from #687) fires before Gpus::init_ep at lib.rs:3434. At this head init_ep is reachable only from the DeepSeek V4 (arch 9, lib.rs:3062) and MiniMax (arch 10, lib.rs:3290) loaders; neither has an artifact on this host. | ep_admission_ornith_tp4_HEAD.json |

| A/B: did master actually serve that arch-6 EP path, i.e. is the inherited refusal a lost working path? | python3 $HW_GATE_EVIDENCE/ep_probe.py $HW_GATE_BASE_BIN/daemon $HIPFIRE_MODELS_DIR/ornith-1.5-35b-a3b.mq4r 4 $HW_GATE_EVIDENCE/ep_admission_ornith_tp4_MASTER.json --devices 0,1,2,3 --generate | Master loads it (16.7 s, '[loader] EP load: tp=4 arch=qwen35 experts=256', vram 32548 -> 25512) but generate fails immediately after gen_start with 'EP arch mismatch (expected dense Qwen TP)'; unload returns vram to 32400. So master's init_tp-based arch-6 EP load was never servable; #687 moved a guaranteed post-load failure to admission. Not a regression of a working path, and not this PR's change. | ep_admission_ornith_tp4_MASTER.json |

| Does the multi-GPU path that IS reachable at this head (dense Qwen3.5 TP=4 through the same load_model_ep_qwen35 function, in the file this PR touches) still load, serve, unload and reload identically to master? | for lane in HEAD:$HW_GATE_BIN MASTER:$HW_GATE_BASE_BIN; do python3 $HW_GATE_EVIDENCE/ep_probe.py $bin/daemon $HIPFIRE_MODELS_DIR/qwen3.6-27b.mq4 4 $HW_GATE_EVIDENCE/dense_tp4_qwen36_27b_cycles_$label.json --devices 0,1,2,3 --generate --cycles 2; done | Both binaries: rank-0 vram_free_mb sequence [32548, 27104, 32320, 27026, 32238] across load/unload/reload/unload — identical to the MB; loads 13.7/14.4 s (HEAD) vs 14.5/14.4 s (master); '[loader] dense qwen TP load complete: 4 ranks'. Decoded reasoning identical on both ('Thinking Process: 1. Identify the core question: The user is asking for the capital of France. 2. Retrieve knowledge: Capital of France = Paris...'); both terminate with the fail-closed 'open think span at end of generation' at my 48-token budget, as designed for a thinking model. | dense_tp4_qwen36_27b_cycles_MASTER.json |

| Does the 4-rank dense TP path produce a complete coherent answer on the PR binary across a reload? | python3 $HW_GATE_EVIDENCE/ep_probe.py $HW_GATE_BIN/daemon $HIPFIRE_MODELS_DIR/qwen3.6-27b.mq4 4 $HW_GATE_EVIDENCE/dense_tp4_qwen36_27b_nothink_HEAD.json --devices 0,1,2,3 --generate --cycles 2 (thinking_enabled=false) | Cycle 0 and cycle 1 (after unload+reload) both decode 'The capital of France is Paris.' (8 tokens); vram sequence again [32548, 27104, 32320, 27026, 32238]. The 'commit_ready -> aborted' terminal is the two-phase wire contract (crates/hipfire-generate/src/dense.rs:3711); my raw driver sends no 'commit', so the daemon's 30 s abort is protocol, not a defect. | dense_tp4_qwen36_27b_nothink_HEAD.json |

| Can WeightTensor::free_all in LlamaWeights::free_gpu / Qwen2Weights::free_gpu double-free a shared sidecar (Sol's suspected regression at llama.rs:686)? | source read: crates/hipfire-runtime/src/llama.rs:526-540, paro.rs:284, hfq.rs:1489/1544, weight_backend.rs:233-247 | No double-free by construction: free_all skips PARO sidecars whose ParoRotation.is_alias is true (alias_paro_rotation sets is_alias: true at paro.rs:284; owners at paro.rs:189 / hfq.rs:1838 set false); awq_scale is a fresh per-weight upload via load_awq_scale at hfq.rs:1489 and :1544 (never a shared handle); the tied lm_head alias (weight_backend.rs tied_lm_head_alias) carries paro: None / awq_scale: None, and free_gpu skips the whole output weight when lm_head_aliases_embd. Static only — no artifact on this host reaches these two free paths. | README.md (citations) |

unproven:

• LlamaWeights::free_gpu sidecar reclamation (llama.rs:685): needs an arch 0/1 artifact carrying AWQ (<w>.awq_scale.weight) or PARO (krot>0) sidecars, cycled load/unload/reload with diag vram_free_mb — master should lose one sidecar set per weight per layer per reload, HEAD should stay flat. The registry's only arch-1 artifacts (qwen3:0.6b, qwen3:8b, HF4, no sidecars) are not on this host, and no registry artifact at all is arch 0/1 with an awq/paro recipe; a locally quantized v3-awq-f1 Qwen3 (arch 1) on any single GPU would prove it.

• Qwen2Weights::free_gpu (qwen2.rs:283): arch 7 (vibethinker:3b / :3b-mq6) is not on this host; by the diff's own note no sidecar is allocated for Qwen2 today, so this is behaviorally a no-op until one is.

• Gpus::init_ep on hardware (multi_gpu.rs:329): at this head only the DeepSeek V4 (deepseek-v4-flash*, arch 9) and MiniMax (minimax-m2.7, arch 10) EP loaders reach it, on a >=2-GPU host; neither artifact is present here. The Ep-axis mesh is pinned only by the unit test device_mesh_ep_group_and_tp_absent, and feat(runtime): add DeviceMesh topology (fixed #673 G1) #681's mesh still has no runtime reader, so a wrong axis would be invisible at runtime either way.

rationale: The hard floor fired on registry/models.json and registry/v1.json, so this decision cannot be merge-staging; it is hold, not block, because every mandatory route passed on both lanes (qwen3.6:27b, ornith-1.5:35b-a3b-mq4r, lfm2.5:1.2b, qwen3.8:27b-mq4-xt batteries and the lfm2.5 chain all coherent; redline capture stable, HIP/PM4 bit-exact on gfx1100 and gfx1201) and nothing I ran regressed. Two facts the maintainer needs: (1) no PR-authored commit touches the registry — git diff 3bd2914b..3f3d0bc6e -- registry/ is empty; the +108/-4 registry delta arrives only through a0aa126 'Merge beta into fix/vram-leak-relands' and is PR #686's already-gated sidecar pairing, so the floor is reacting to staging content, not to this PR; (2) HEAD a0aa126 is already an ancestor of origin/beta (7df822e, 21:30 UTC, previous round's merge-staging), so this is a re-gate of a merged head — the class of run PR #712 exists to stop. What would clear the floor without any code change: compute policy_paths in select.py against the PR's merge-base with the staging branch (or drop files whose blob equals origin/beta's) rather than against master. On the PR's own claims: the previous round's ornith 4x EP evidence for init_ep is no longer valid at this head — ep_admission_ornith_tp4_HEAD.json shows #687's qwen35_ep_moe_refusal (loader/lib.rs:3416) refusing the load in 0.28 s before Gpus::init_ep (lib.rs:3434), and ep_admission_ornith_tp4_MASTER.json shows master loaded that topology but could not serve it ('EP arch mismatch (expected dense Qwen TP)' on generate), so the inherited refusal removed no working path and init_ep is now reachable only from the arch-9/arch-10 loaders, which have no fixture here. The reachable multi-GPU path in the touched function — qwen3.6:27b dense TP=4 — is byte-identical between HEAD and master across two load/generate/unload cycles (vram [32548,27104,32320,27026,32238] on both; identical decoded text; thinking-off run answers 'The capital of France is Paris.' both cycles). The sidecar-free change is sound by ownership (free_all honors ParoRotation.is_alias at paro.rs:284; awq_scale is per-weight at hfq.rs:1489/1544; the tied lm_head alias carries no sidecars) but is unproven on hardware because no arch 0/1/7 artifact exists on this host, and the registry has no AWQ/PARO llama-family artifact at all. I would put my name on the code; I cannot put it on the sidecar claim as measured, and the floor forbids merging from this seat regardless.

hard floor: ['policy_paths: registry/models.json,registry/v1.json']

soft floor: ["coverage_gaps: ['No repeated AWQ/PARO Llama-family unload/reload VRAM measurement exercised WeightTensor::free_all sidecar reclamation.', 'No multi-rank EP load exercised Gpus::init_ep or inspected Ep/Tp mesh dimensions.', 'No Qwen2 fixture exercised its changed unload path.', 'No DFlash run exercised sidecar resolution, mandatory-draft failure, context exhaustion, emitter rollback, or speculative terminal behavior.', 'No Gemma4 lowered/MoE artifact exercised the new admission refusal.', 'CLI pull/rm behavior and registry sidecar policy were not exercised end to end.', 'The changed registry/models.json and registry/v1.json policy requires human review.']", 'model needs-human']

[hipfire-fable]

Merging to staging. I took the EP change onto real hardware: a 4x gfx1201 expert-parallel load of ornith-1.5:35b-a3b-mq4r goes through Gpus::init_ep, completes on 4 ranks, and unloads/reloads with VRAM readings byte-identical to the master binary; the engine's 'EP arch mismatch' refusal at generate is the pre-existing #683 behavior and shows identically on master. The sidecar-free change could not run on LlamaWeights here (no llama-arch artifact in the registry on the gate hosts), so I verified at every sidecar construction site that nothing is shared between weights, and exercised the same free_all body four times in the daemon unload path on an AWQ artifact with flat VRAM and identical output after each reload. If you have an AWQ or PARO llama/qwen3-dense artifact on hand, a load/generate/unload loop watching vram_free_mb would close the last gap. The diff is 105 lines across three behaviors. (1) Gpus::init_ep (multi_gpu.rs:329-366) is field-for-field init_tp (multi_gpu.rs:280-317) except mesh: DeviceMesh::rect(Ep, n); grep of crates/ shows the mesh field is read only by constructors and tests, so the runtime delta is zero. I exercised it on real hardware: a 4x gfx1201 EP load of ornith-1.5:35b-a3b-mq4r on the PR daemon logged 'EP load: tp=4 arch=qwen35' after init_ep, completed on 4 ranks, unloaded, reloaded and unloaded again with rank-0 vram_free_mb 32548/25512/32400/25510/32398 -- the identical five numbers the master daemon produced from the same stdin. The harness battery at tp=4 returned 5 empty turns on both binaries because the engine refuses Qwen3.5-MoE EP generation with 'EP arch mismatch (expected dense Qwen TP)' (qwen.rs:385) -- pre-existing on master, the #683 family that #687 refuses at admission; not this PR's regression, and the two serve logs differ only in timing lines. (2) LlamaWeights::free_gpu -> WeightTensor::free_all: no llama-arch fixture exists on this host so the function itself did not run; the risk it introduces is a double free of a shared sidecar, and I checked every sidecar construction site (hfq.rs:1401/1489/1544, paro.rs:183-191, paro.rs:264-286 with the is_alias skip at llama.rs:530, weight_backend.rs:245-246 with the preserved lm_head_aliases_embd guard at llama.rs:688): each sidecar is uniquely owned by one WeightTensor. The free_all body itself ran 4x in the daemon's real unload path on an AWQ-sidecar artifact (qwen3.5-0.8b.mq4, 138 awq_scale tensors) with vram_free_mb flat at 32314 after every unload and byte-identical decoded output after every reload. (3) Qwen2Weights::free_gpu: no sidecar is ever constructed for Qwen2 (qwen2.rs:640, 687 are the only sites, both None), so free_all is exactly the old free_tensor(.buf). The mandatory batteries (qwen3.6:27b, qwen3.8:27b-mq4-xt, ornith-1.5:35b-a3b-mq4r, lfm2.5:1.2b on gfx1100 and gfx1201) all decoded coherently; none of them touches a changed line, which is why Sol's coverage read is correct and mine adds the multi-GPU and unload-cycle routes above. One observation outside this diff for the maintainer: the gfx1100 lfm2.5:1.2b chain turn 2 rendered subscript identifiers as 'Distance��' while the gfx1201 lane rendered the same turn cleanly; the PR touches neither LFM2 nor tokenizer decode and I have no gfx1100 device in this sandbox to A/B it, so it is noted, not charged. No regression is plausible against master: the new code paths either did not change runtime behavior (mesh, Qwen2) or only free buffers that were previously leaked and are provably uniquely owned. I would put my name on this merge to staging. Override Sol needs-human: I closed the multi-GPU gap myself on this 5x gfx1201 host: Gpus::init_ep was entered through load_model_ep_qwen35 (ornith-1.5:35b-a3b-mq4r, tp=4), the 4-rank load/unload/reload cycle is byte-identical to the master binary in every VRAM reading and log line, and DeviceMesh has no runtime reader (grep: only the three constructors and tests), so the only behavioral delta is the recorded axis pinned by the new unit test. The remaining gaps (LlamaWeights::free_gpu on a real AWQ/PARO llama-arch artifact; Qwen2) cannot be exercised on any hw-gate host with the current registry, and I verified at every sidecar construction site that no sidecar is shared between two WeightTensors, so the only new failure mode (double free) is excluded structurally; the free_all primitive itself ran 4x in the daemon unload path on an AWQ-sidecar artifact with flat VRAM.

[Kaden-Schutt]

Merged to beta as 7df822e9b (branch tip a0aa12681).

Verdict was merge-staging; the seat was stopped only by the recurring generated-map 409 (staging_merge_conflict on crates/hipfire-runtime/map.md). Merged beta in, regenerated the block, merged.

The decide phase measured both claims on the 5×gfx1201 host rather than reasoning about them:

• 4× EP load of ornith-1.5:35b-a3b-mq4r on the PR daemon logged EP load: tp=4 arch=qwen35 after init_ep, completed on 4 ranks, then unload → reload → unload with rank-0 vram_free_mb 32548 / 25512 / 32400 / 25510 / 32398 — the identical five numbers the master daemon produced from the same stdin. That is the evidence the Ep mesh axis is a zero-runtime-delta rename.
• Gpus::init_ep is field-for-field init_tp except mesh: DeviceMesh::rect(Ep, n), and the mesh field is read only by constructors and tests.

Sol's needs-human was coverage, not a regression, and it is worth writing down as follow-up rather than losing: no AWQ/PARO Llama artifact was cycled with free-VRAM diagnostics (so the sidecar-leak fix itself is argued, not measured), and the Qwen2 free_all path has no fixture on either lane. Both are recorded as unproven in the decision artifact.

21 tests pass in hipfire-runtime + hipfire-loader at the merged tip.

[hipfire-sol]

hw-gate sol verdict needs-human: Hardware evidence is clean for ordinary load/serve and kernel execution: both lanes passed, Qwen3.5 capture sequences were stable, and HIP/PM4 state was bit-exact. It does not prove either principal PR claim: crates/hipfire-runtime/src/llama.rs:686 was not tested with repeated sidecar-bearing unload/reload VRAM accounting, and crates/hipfire-runtime/src/multi_gpu.rs:319 was not exercised with multiple EP ranks. The much broader diff also changes DFlash state handling and registry policy without corresponding routes, so human review is required.

[hipfire-fable]

Holding on the hard floor, not on your code: the registry/models.json and registry/v1.json edits that tripped the policy rule come in through your 'Merge beta' commit from PR #686, not from any commit you authored, and this head is already on beta from the previous round — the gate re-ran a merged PR. I re-checked the two claims on the 5x gfx1201 host: at this head ornith 4x EP is refused by #687's admission before init_ep ever runs (and master, while it loaded that topology, could not generate on it), so init_ep is now only reachable from DeepSeek V4 / MiniMax loads, which this host cannot run; the reachable 4-rank dense TP path for qwen3.6:27b is byte-identical to master across load/generate/unload/reload. The sidecar free_all change reads correct by ownership (alias guard, per-weight AWQ upload), but there is no AWQ/PARO llama-family or Qwen2 artifact here to measure it, so it stays recorded as unproven rather than proven. Evidence is under fable-evidence/ with a README index. The hard floor fired on registry/models.json and registry/v1.json, so this decision cannot be merge-staging; it is hold, not block, because every mandatory route passed on both lanes (qwen3.6:27b, ornith-1.5:35b-a3b-mq4r, lfm2.5:1.2b, qwen3.8:27b-mq4-xt batteries and the lfm2.5 chain all coherent; redline capture stable, HIP/PM4 bit-exact on gfx1100 and gfx1201) and nothing I ran regressed. Two facts the maintainer needs: (1) no PR-authored commit touches the registry — git diff 3bd2914b..3f3d0bc6e -- registry/ is empty; the +108/-4 registry delta arrives only through a0aa126 'Merge beta into fix/vram-leak-relands' and is PR #686's already-gated sidecar pairing, so the floor is reacting to staging content, not to this PR; (2) HEAD a0aa126 is already an ancestor of origin/beta (7df822e, 21:30 UTC, previous round's merge-staging), so this is a re-gate of a merged head — the class of run PR #712 exists to stop. What would clear the floor without any code change: compute policy_paths in select.py against the PR's merge-base with the staging branch (or drop files whose blob equals origin/beta's) rather than against master. On the PR's own claims: the previous round's ornith 4x EP evidence for init_ep is no longer valid at this head — ep_admission_ornith_tp4_HEAD.json shows #687's qwen35_ep_moe_refusal (loader/lib.rs:3416) refusing the load in 0.28 s before Gpus::init_ep (lib.rs:3434), and ep_admission_ornith_tp4_MASTER.json shows master loaded that topology but could not serve it ('EP arch mismatch (expected dense Qwen TP)' on generate), so the inherited refusal removed no working path and init_ep is now reachable only from the arch-9/arch-10 loaders, which have no fixture here. The reachable multi-GPU path in the touched function — qwen3.6:27b dense TP=4 — is byte-identical between HEAD and master across two load/generate/unload cycles (vram [32548,27104,32320,27026,32238] on both; identical decoded text; thinking-off run answers 'The capital of France is Paris.' both cycles). The sidecar-free change is sound by ownership (free_all honors ParoRotation.is_alias at paro.rs:284; awq_scale is per-weight at hfq.rs:1489/1544; the tied lm_head alias carries no sidecars) but is unproven on hardware because no arch 0/1/7 artifact exists on this host, and the registry has no AWQ/PARO llama-family artifact at all. I would put my name on the code; I cannot put it on the sidecar claim as measured, and the floor forbids merging from this seat regardless.