fix(arch-llama): unify Site A attention dispatch + expand llama HFQ KV-mode policy

[fivetide]

Summary

Deferred items from PR #463 (transparent loading). Three commits:

1. fix(arch-llama): unify Site A attention dispatch (arch.rs:242)

Llama::forward_scratch_layers in hipfire-arch-llama had a hardcoded Q8-shaped KvTierInputs literal (quant_asym*: false, v_mode_bits: 8, givens_cos/sin: None). For asym or HFQ8 KV caches this would either panic or silently select the wrong kernel.

Changes in crates/hipfire-arch-llama/src/arch.rs:

• Replace the literal with ..kv_cache.tier_inputs() struct update (all quantization flags now come from cache state)
• Thread givens_cos/sin from the cache into AttnParams (asym attend kernels unwrap these)
• Pass flash_partials: Some(&scratch.attn_partials) (required for flash attend tiles)
• Add conditional (k_scales, v_scales) for HFQ8 (attend kernel panics on None for that tier)
• Use gpu.graphs.capture_mode instead of hardcoded false

Byte-identical no-op for Q8 caches.

2. test(dispatch-tests): KV-tier dispatch sweep

9 GPU-free tier-routing tests in hipfire-dispatch-tests/src/llama.rs covering Q8, asym2/3/4, FWHT+asym4, HFQ4, HFQ8, Q4, F32. Each verifies uses_givens and the paired write_key/attend_key.

3. feat(kv-mode): expand LLAMA_HFQ_POLICY + wire kv_mode_override

• kv_mode.rs: LLAMA_HFQ_POLICY now accepts [Q8, Asym3, Asym4] (was [Q8] only). Asym2/Fwht* have no Flat KV constructor and silently fall to Q8.
• llama.rs::KvCache::from_mode: new clean-error gate for Asym3 alongside the existing Asym4 gate — at head_dim≠256 returns a user-readable error instead of panicking in new_gpu_asym3_capped.
• carrier.rs::load_bundle: was hardcoding "" for kv_mode_override, silently ignoring explicit --kv-mode requests for HFQ llama loads since the carrier was introduced. Now passes ctx.kv_mode_override.

Validated sweep on qwen3-0.6b-llama.mq4 (head_dim=128, gfx1151):

--kv-mode	Result
q8	LOADED, coherent output
asym3	Clean error: "asym3 KV cache requires head_dim=256"
asym4	Clean error: "asym4 KV cache requires head_dim=256"
asym2	LOADED → Q8 (silently downgraded, no Flat constructor)

Test plan

• cargo test -p hipfire-dispatch-tests — 9 new tier tests pass
• cargo test -p hipfire-runtime --lib -- kv_mode — truth-table tests pass
• Q8 coherence: qwen3-0.6b-llama.mq4 with --kv-mode q8 produces coherent output
• Asym gate: --kv-mode asym3 on head_dim=128 model returns clean load error

🤖 Generated with Claude Code

[fivetide]

Rebased onto current master (d44f89e7), conflict resolved.

Conflict was entirely stale-base noise. The branch forked at 2026-06-21, before the large KV/dispatch churn from the merged speculator-abstraction PR (#477), so it carried duplicate loader/tui changes that now collide. The PR's real work is just 3 commits / 5 files, none of which touch the conflicting files — so I cherry-picked the 3 real commits onto fresh master (all applied cleanly) and dropped the 4 trailing style: rustfmt/clippy commits and the stale base. Net diff: 214 insertions / 45 deletions across arch.rs, carrier.rs, kv_mode.rs, llama.rs, and the new dispatch-tests sweep.

Still needed on master: I confirmed master's Site A is unfixed — crates/hipfire-arch-llama/src/arch.rs still has givens_cos/sin: None and the hardcoded Q8-shaped KvTierInputs literal. This PR replaces that with kv_cache.tier_inputs() + threaded givens/scales/flash_partials, which is byte-identical for Q8 caches and only changes the asym/HFQ8 paths (which currently panic).

Validation (CPU):

• New Site-A KV-tier dispatch sweep (added by this PR): 16/16 pass — incl. llama_kv_tier_asym{2,3,4}_uses_givens, asym4_fwht_uses_givens (the fix), and q8_uses_q8_kernels_no_givens (byte-identical common path).
• kv_mode truth-table tests: 7/7 pass, incl. truth_table_llama_hfq_expanded (the asym3/asym4 LLAMA_HFQ_POLICY change).
• Full workspace build (lib + bins + examples): clean.

Scope caveat: validation is the targeted dispatch/policy unit sweep + byte-identical-Q8 design — I did not run an end-to-end GPU coherence pass with asym KV on a llama-Dir model this round (the path the fix actually enables). Recommend that as the pre-merge gate, since it's the surface my notes previously flagged as panic/garbage-prone. The Q8 default path is byte-identical and unit-confirmed, so the common case carries no regression risk.

Now MERGEABLE.

[fivetide]

✅ Validated in a local integration stack (#497→#473→#495→#468→#466→#465) on gfx1151 — cargo build --release --workspace --locked green; scripts/coherence-gate.sh correctness PASS (no hard errors) and scripts/qwen2-reset-gate.sh PASS (0 ceiling resets, 0 panics). Merges clean onto current master with no conflicts. (The pflash perf stage false-fails on this box: gfx1100 baseline vs gfx1151 — untouched path.)

[Kaden-Schutt]

Folded into #534 as a cleanly rebased/cherry-picked delta on beta for v0.3.0, preserving the newer Redline and modular-runtime line. Closing this source PR in favor of the consolidated release PR. Thank you for the contribution.