fix(hw-gate): drive DFlash in the gate; stop refusing user-path CLI diffs

[Kaden-Schutt]

Two policy gaps this ladder exposed. Both are gate plumbing, no product code.

1 · The gate has never run DFlash

#686 (draft sidecars), #691 (draft ctor rollback), #692 (primer replay) and #702 (dedicated verify kernels) all went through this gate with every lane green while speculation never once executed. #692's DFlash-arm defect — primer replay systematically missing the most recent assistant body — surfaced only because a seat thought to drive twenty turns by hand. That is not a gate.

The load bucket now runs battery-dflash, the serve bucket chain-dflash: the same prompts with --dflash on and an explicit --draft.

• on, not auto — auto silently falls back to AR when the draft is missing, and a route that can pass without speculating proves nothing.
• explicit --draft — the canonical xt trunk is a symlink out of the models dir, so the daemon's filename auto-match finds nothing and would quietly run AR.
• dflash_draft is a candidate list, because the lanes hold different drafts: hiptrx has qwen36-27b-dflash-mq4.hfq and no qwen38, hipx has qwen38-27b-dflash-mq4.hfq and no qwen36.

skip is neither pass nor fail

This is the part worth reviewing. The aggregation was all(status == "pass"), so a skipped mode would have counted as a fixture failure — a false negative on evidence the host never had — while treating it as a pass would claim coverage that did not happen. Skips are now recorded, reported in the reason string, and excluded from the verdict; a genuine failure alongside a skip still fails.

Coverage is asymmetric until both hosts hold both drafts. Pulling qwen38-27b-dflash-mq4.hfq to hiptrx and qwen36-27b-dflash-mq4.hfq to hipx (0.92 GB each) makes it symmetric — that is a disk decision, so this reports skip rather than silently pulling ~2 GB onto your machines.

2 · Sol refused hardware for user-path CLI diffs

The rule "filesystem access beyond model/cache/temp paths" caught #689 for adding --prompt-file to hipfire bench, and cost that rung a hardware lane until hw-run overrode it. hipfire is a CLI inference engine: users name models, prompts, drafts and sidecars at invocation, and the gate's own harness passes exactly those flags.

sol.md now separates whose path it is. An explicit argument is ordinary product work; credentials, dotfiles, SSH or cloud config, /proc//sys beyond device enumeration, traversal assembled from something other than an argument, or a read whose result leaves the process still warrant refusal.

Tests

Eight new cases: flag translation (battery-dflash → --mode battery --dflash on --draft …), plain battery never receiving a draft, per-lane draft selection, skip-not-fail with the harness never invoked, chain-dflash keeping its own prompts, the skip-vs-genuine-failure aggregation, and a manifest assertion that the buckets carry the routes. 113/113 hw-gate tests pass.

Policy-floor by construction, so it cannot self-merge.

[hipfire-sol]

hw-gate sol prelim

summary: Changes hardware-gate policy and orchestration: load/serve buckets gain DFlash variants, those variants translate to existing battery/chain harness modes with forced speculation and an explicit locally selected draft, absent draft candidates produce recorded skips, and skipped modes are excluded from fixture aggregation. It also narrows Sol's filesystem-refusal policy for explicit user-supplied CLI paths.

run_hardware: true
run_hardware_reasons: The diff is safe to execute: it invokes the existing checked-in serve harness, reads only declared local model/draft fixtures and gate prompt files, and adds no network, credential, dependency, toolchain, unsafe, or opaque behavior.; The harness changes are understandable and bounded, but real execution is needed to determine whether cross-version draft pairing loads and genuinely speculates rather than failing or silently losing coverage.

routes:

mode	tag	source	why
battery	qwen3.6:27b	sol	Exercise the changed load-oriented harness path on the canonical dense target; hardware evidence must show whether the available draft pairing is compatible and coherent.
chain	qwen3.8:27b-mq4-xt	sol	Exercise multi-turn state behavior on the xt target whose explicit draft selection motivated the change.

unavailable_routes:

(none)

claim_assessment: The author claims the new routes force DFlash, choose a lane-local draft, and preserve a neutral skip state. Unit-test counts do not prove that. Evidence must show the selected draft is model-compatible, the daemon actually enters speculation, decoded battery/chain turns remain coherent, and a lane with no draft is reported as uncovered without allowing required DFlash coverage to disappear from the overall gate verdict.

questions_for_author:

• What establishes that qwen36 and qwen38 DFlash drafts are valid interchangeable partners for both target fixtures?
• Why does the implementation skip when all explicitly declared draft candidates are absent despite the code comment and manifest invariant saying a missing declared artifact remains a hard failure?

[hipfire-sol]

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

lane hiptrx (gfx1201)

hw-gate evidence

field	value
base	d1da3eae2835e3b87134f2747f3d01e9978e4fb9
head	744aeb3b3eaedc495a6d8e62b8b769c6630dee88
buckets
host gfx	gfx1201
host rocm	7.15.26333-0000000
device	1
runner	hiptrx
daemon_md5	17ba7dfab49aa1c1110d3f412f072648
hipfire_md5	cb62380a15b0e3481fad7fdfb1ae04ee
build_seconds	42.72876310348511
verdict	pass
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

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

battery — exit 0 seconds 31.3 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	3.1	15.2	False	False	False	True
battery	1		stop	30	0	83	72	504.5	36.0	False	False	False	True
battery	2		stop	31	0	19	16	521.9	36.2	False	False	False	True
battery	3		stop	47	0	31	8	565.0	36.1	False	False	False	True
battery	4		stop	47	0	17	7	563.7	36.1	False	False	False	True

qwen3.6:27b battery turn 0

```python
def longest_substring_without_repeating(s: str) -> str:
    char_index = {}
    start = 0
    max_len = 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_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 scatters off air molecules in all directions, with shorter blue wavelengths dispersing more effectively to make the sky appear blue. As the sun sets, its light travels through a thicker layer of the atmosphere, causing most of the blue light to scatter away before reaching your eyes. This leaves the longer red and orange wavelengths to dominate the visual field, resulting in the characteristic red 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

17 + 26 = 43

Answer: 43

qwen3.8:27b-mq4-xt

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

chain — exit 0 seconds 35.3 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	44	0	120	51	11.3	13.6	False	False	False	True
chain	1		stop	224	0	209	115	102.4	65.5	False	False	False	True
chain	2		stop	463	0	82	67	191.1	36.3	False	False	False	True
chain	3		stop	583	0	121	95	722.3	35.4	False	False	False	True
chain	4		stop	740	0	72	55	802.5	31.5	False	False	False	True

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

```python
def merge_sorted(a, b):
    """Merge two sorted lists into a single sorted list."""
    result = []
    i, j = 0, 0
    while i < len(a) and j < len(b):
        if a[i] <= b[j]:
            result.append(a[i])
            i += 1
        else:
            result.append(b[j])
            j += 1
    # Append remaining elements
    result.extend(a[i:])
    result.extend(b[j:])
    return result
```

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

To find the total distance traveled, we calculate the distance for each segment of the trip and then add them together. The formula for distance is:

$$ \text{Distance} = \text{Speed} \times \text{Time} $$

**Step 1: Calculate the distance for the first segment**
*   Speed = 60 mph
*   Time = 2.5 hours
*   Distance = $60 \times 2.5 = 150$ miles

**Step 2: Calculate the distance for the second segment**
*   Speed = 40 mph
*   Time = 1.5 hours
*   Distance = $40 \times 1.5 = 60$ miles

**Step 3: Add the distances together**
*   Total Distance = $150 + 60 = 210$ miles

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

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

The seasons are caused by the tilt of Earth's axis relative to its orbital plane, not by the planet's distance from the Sun. As Earth orbits the Sun, this axial tilt causes specific hemispheres to be angled toward or away from the Sun, changing the intensity and duration of sunlight they receive. This variation in solar exposure drives the cyclical patterns of spring, summer, autumn, and winter.

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

Elias wiped the salt spray from his goggles and stared at the sleek, silver pod resting on the jagged rocks, far from any known shipping lane. He approached cautiously, his boots crunching on the wet stone, until he noticed the faint, rhythmic pulsing of light emanating from the object's seam. When he pried it open with his crowbar, he didn't find cargo or a message, but a single, perfect glass marble containing a swirling, miniature galaxy. He held it up to the dawn, realizing his lonely watch was not over, but had just truly begun.

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

1. Use descriptive variable and function names that clearly convey their purpose.
2. Keep functions small and focused on a single responsibility.
3. Write comprehensive unit tests to verify behavior and catch regressions.
4. Maintain consistent code style and formatting throughout the project.
5. Add concise comments only to explain complex logic, not obvious syntax.

kernel

not run

lane hipx (gfx1100)

hw-gate evidence

field	value
base	d1da3eae2835e3b87134f2747f3d01e9978e4fb9
head	744aeb3b3eaedc495a6d8e62b8b769c6630dee88
buckets
host gfx	gfx1100
host rocm	7.15.26333-0000000
device	0
runner	hipx
daemon_md5	17ba7dfab49aa1c1110d3f412f072648
hipfire_md5	cb62380a15b0e3481fad7fdfb1ae04ee
build_seconds	49.09681797027588
verdict	pass
logs_dir	hw-gate-logs

fixtures

qwen3.6:27b

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

battery — exit 0 seconds 28.5 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	146	66	3.3	18.8	False	False	False	True
battery	1		stop	30	0	79	68	437.1	50.3	False	False	False	True
battery	2		stop	31	0	19	16	450.5	50.4	False	False	False	True
battery	3		stop	47	0	31	8	464.9	50.3	False	False	False	True
battery	4		stop	47	0	6	2	463.7	50.2	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
        current_length = end - start + 1
        if current_length > max_length:
            max_length = current_length
            max_start = start
            
    return s[max_start:max_start + max_length]
```

qwen3.6:27b battery turn 1

During the day, sunlight scatters through the atmosphere, and shorter blue wavelengths are dispersed more effectively than longer red wavelengths, making the sky appear blue. At sunset, sunlight travels through a much thicker layer of the atmosphere to reach your eyes, causing most of the blue light to scatter away completely. This leaves the longer red and orange wavelengths to dominate the sky's appearance as the sun sets.

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

qwen3.8:27b-mq4-xt

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

chain — exit 0 seconds 42.3 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	44	0	112	46	16.0	28.1	False	False	False	True
chain	1		stop	216	44	182	90	60.5	48.9	False	False	False	True
chain	2		stop	428	44	75	61	186.9	48.8	False	False	False	True
chain	3		stop	541	44	110	87	278.4	49.2	False	False	False	True
chain	4		stop	687	44	73	55	871.4	49.0	False	False	False	True

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

```python
def merge_sorted(a, b):
    """Merge two already-sorted lists into one sorted list."""
    result = []
    i = j = 0
    while i < len(a) and j < len(b):
        if a[i] <= b[j]:
            result.append(a[i])
            i += 1
        else:
            result.append(b[j])
            j += 1
    result.extend(a[i:])
    result.extend(b[j:])
    return result
```

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

To find the total distance, calculate the distance for each segment separately using the formula:

$$\text{Distance} = \text{Speed} \times \text{Time}$$

**Step 1: Distance for the first segment**
*   Speed = 60 mph
*   Time = 2.5 hours
*   Distance = $60 \times 2.5 = 150$ miles

**Step 2: Distance for the second segment**
*   Speed = 40 mph
*   Time = 1.5 hours
*   Distance = $40 \times 1.5 = 60$ miles

**Step 3: Total distance**
*   Total Distance = $150 + 60 = 210$ miles

**Final Answer:** 210 miles

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

The seasons are caused by the Earth's axial tilt of approximately 23.5 degrees relative to its orbital plane around the Sun. As Earth orbits the Sun, this tilt causes different hemispheres to receive varying amounts of direct sunlight and daylight hours throughout the year. When a hemisphere is tilted toward the Sun, it experiences summer, while the opposite hemisphere experiences winter.

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

Elias wiped the salt from his eyes, expecting nothing but kelp and debris when he stepped out to check the lower rocks at dawn. Instead, he found a sealed brass canister, tarnished by the sea, wedged tightly between two jagged stones. His fingers trembled as he cracked the wax seal, revealing a map drawn in ink that smelled faintly of lavender, not brine. The coordinates on the parchment pointed directly to the small, unmarked island just beyond his horizon, where a light had never before been lit.

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

1. Use descriptive variable and function names to make intent clear without comments.
2. Keep functions small and focused on a single responsibility.
3. Write unit tests for all new logic to prevent regressions.
4. Follow established coding standards and style guides consistently.
5. Add concise comments only to explain complex logic or non-obvious decisions.

kernel

not run

[hipfire-sol]

hw-gate sol verdict

{
  "claim_verdict": "not-exercised",
  "confidence": 0.98,
  "coverage": {
    "gaps": [
      "The evidence contains only plain battery and chain modes; neither battery-dflash nor chain-dflash ran.",
      "No explicit draft was selected, loaded, or shown to enter speculation.",
      "Cross-version qwen3.6/qwen3.8 draft compatibility was not exercised.",
      "The skip aggregation path was not exercised on hardware, and run.py:542 still permits all declared draft candidates to be absent without failing the fixture.",
      "The changed Sol filesystem policy is a policy-file change requiring human review."
    ],
    "surfaces_evidenced": [
      "load",
      "serve"
    ],
    "surfaces_touched": [
      "policy",
      "load",
      "serve",
      "speculative-decode",
      "harness"
    ]
  },
  "decision": "needs-human",
  "eyeball": [
    "All five qwen3.6 battery responses on gfx1201 and gfx1100 are coherent, answer their prompts, contain required substrings, and show no attractor, empty output, runaway, or special-token leakage.",
    "All five qwen3.8 chain responses on gfx1201 and gfx1100 are coherent; the merge function, 210-mile calculation, seasons explanation, lighthouse story, and coding-practices list answer their prompts without degeneration.",
    "A human should inspect scripts/hw-gate/run.py:542-548 because absence of every explicitly declared draft produces skip and is excluded from the verdict, allowing a selected DFlash route to contribute no evidence."
  ],
  "phase": "verdict",
  "rationale": "The ordinary qwen3.6 battery and qwen3.8 chain routes passed coherently on both gfx1201 and gfx1100, but hw-gate.json contains no *-dflash mode at all. It therefore proves only the existing AR load/serve paths, not the PR's central forced-DFlash routing, lane-local draft selection, cross-version pairing, or skip semantics. Because scripts/hw-gate/fixtures.json, run.py, and sol.md change gate policy and the relevant behavior was not exercised, this requires human review rather than a greenlight.",
  "regressions": []
}

Floor: hard=['policy_paths: scripts/hw-gate/fixtures.json,scripts/hw-gate/run.py,scripts/hw-gate/sol.md,scripts/hw-gate/tests/test_run.py'] soft=["coverage_gaps: ['The evidence contains only plain battery and chain modes; neither battery-dflash nor chain-dflash ran.', 'No explicit draft was selected, loaded, or shown to enter speculation.', 'Cross-version qwen3.6/qwen3.8 draft compatibility was not exercised.', 'The skip aggregation path was not exercised on hardware, and run.py:542 still permits all declared draft candidates to be absent without failing the fixture.', 'The changed Sol filesystem policy is a policy-file change requiring human review.']", 'model needs-human'] model_decision=needs-human final=needs-human

[hipfire-sol]

hw-gate sol verdict needs-human: The hardware run succeeded only as a build: buckets and fixtures are empty, so it provides no evidence for the new DFlash routes or skip semantics. The diff is a policy-floor change and therefore requires human review. In scripts/hw-gate/run.py:621-623, removing every skipped result makes all([]) true, allowing an all-skipped fixture to pass; scripts/hw-gate/tests/test_run.py:971 tests a copied expression rather than _run_fixture_harness, so it does not close that gap.

[hipfire-sol]

hw-gate sol verdict needs-human: The ordinary qwen3.6 battery and qwen3.8 chain routes passed coherently on both gfx1201 and gfx1100, but hw-gate.json contains no *-dflash mode at all. It therefore proves only the existing AR load/serve paths, not the PR's central forced-DFlash routing, lane-local draft selection, cross-version pairing, or skip semantics. Because scripts/hw-gate/fixtures.json, run.py, and sol.md change gate policy and the relevant behavior was not exercised, this requires human review rather than a greenlight.