docs(agents): compress the load-bearing lessons, repoint their owner links

Four bullets in Key Lessons Learned carried narrative a vault page owns: the
2026-09-20 board swap ([[Dual-R9700-Cooperation]]), the `--override-tensor`
gate ([[Llama Cpp Multi-Device Splits]]), the x4 link budget
([[x4-pair-experiment]]), and the silent-RAM incident plus the Prometheus
`--metrics` law ([[frank-archon]], [[frank-zealot]]). Each keeps its law, its
numbers and its error strings; the causal narration that duplicated a page is
what went. The RAM bullet linked to [[log]] — the incident is owned by
[[frank-archon]], so the link now names the owner.

AGENTS.md

@@ -485,13 +485,12 @@ linked runbooks.
   6 set. AMD desktop platforms put one card on CPU root ports (atomics work) and the rest behind the
   chipset (usually not) — that asymmetry is what blocked TP=2 here. A one-card collective proves
   nothing: NCCL short-circuits at `world_size=1`. [[PCIe Link and Atomics]]
-- **That wall was physical, and a board swap removed it on 2026-09-20.** dankserver took the path of
+- **That wall was physical, and a board swap removed it on 2026-09-20.** dankserver took a board
-  a board whose CPU lanes bifurcate to x8/x8 Gen5, both cards on root ports, every bridge `Routing+`
+  whose CPU lanes bifurcate to x8/x8 Gen5: both cards on root ports, every bridge `Routing+`, BOTH
-  and BOTH endpoints `ReqEn+` — the exact precondition the old test plan named. Bandwidth was never
+  endpoints `ReqEn+`. Bandwidth was never the constraint (a 27B TP=2 allreduce at batch 4 is ≈10 MB
-  the constraint (a 27B TP=2 allreduce at batch 4 is ≈10 MB per token-step, under 0.4 ms at Gen5 x8).
+  per token-step, under 0.4 ms at Gen5 x8); the collective choice is — TheRock's RCCL builds lack
-  The collective choice matters: TheRock's RCCL builds lack gfx1201 device code
+  gfx1201 device code (`invalid device function`), which is why vLLM runs its R4D one-shot allreduce
-  (`invalid device function`), which is why vLLM runs its R4D one-shot allreduce here at all.
+  here at all. Evidence matrix and the three-arm verdict: [[Dual-R9700-Cooperation]].
-  Evidence matrix and the measured three-arm verdict: [[Dual-R9700-Cooperation]].
 - **When a GPU collective dies, read the runtime, not the framework.** The message that identified
   the cause came from `AMD_LOG_LEVEL=3` (`Pcie atomics not enabled, hostcall not supported` →
   `AQL dispatch failed!`), never from vLLM or RCCL — the env var costs one line and prints the
@@ -499,26 +498,23 @@ linked runbooks.
 - **Layer split is the no-collective path, on one vendor's cards and across two — and
   `--override-tensor` is what cannot ride with it.** Layers go in order and only the boundary
   activation crosses PCIe once per forward pass, so it works wherever a collective cannot be
-  dispatched (and on an RCCL-less fork it beats tensor mode, which falls back to a butterfly
+  dispatched (on an RCCL-less fork it beats tensor mode, which falls back to a butterfly allreduce),
-  allreduce); it is also the cross-vendor answer, where a CUDA + Vulkan split held a real
+  and it is the cross-vendor answer too (a CUDA + Vulkan split held a real two-device graph). But
-  two-device graph. But `--override-tensor <group>=<device>` sets `has_tensor_overrides()`, a
+  `--override-tensor <group>=<device>` sets `has_tensor_overrides()`, a REQUIRED term of the
-  REQUIRED term of the scheduler's `pipeline_parallel` flag (llama-context.cpp), and that flag sets
+  scheduler's `pipeline_parallel` flag, and that flag sets `n_copies`: with `n_copies == 1` a
-  `n_copies` in `ggml_backend_sched_new`: with `n_copies == 1` a pre-allocated weight on a
+  pre-allocated weight on a non-primary device has no backend and the graph aborts (reported with
-  non-primary device has no backend and the graph aborts (ggml-backend.cpp, reported with the leaf's
+  the leaf's op name `(NONE)`). `=CPU` is the only override target that works — the CPU backend is
-  own op name `(NONE)`). `=CPU` is the only override target that works, because the CPU backend is
+  last in the list — and pure Vulkan changes nothing: the gate is the override, not the driver mix.
-  last in the list, and pure Vulkan changes nothing — the gate is the override, not the driver mix.
+  Pin the spec-decode drafter to one device; on a split its home is the card holding the target's
-  Pin the spec-decode drafter to one device: an unpinned draft under tensor mode aborts the KV build,
+  last layers. [[Llama Cpp Multi-Device Splits]] [[x4-pair-experiment]]
-  and on a split its home is the card holding the target's last layers.
-  [[Llama Cpp Multi-Device Splits]] [[x4-pair-experiment]]
 - **A host↔3090 link at x4 measures 6.72 GB/s, and that is a design constraint, not a detail.**
-  Pinned H2D 6.72 / D2H 6.60 GB/s (≈85 % of the Gen4 x4 ceiling), async no better than sync and
+  Pinned H2D 6.72 / D2H 6.60 GB/s (≈85 % of the Gen4 x4 ceiling), pageable within 1 % of pinned —
-  pageable within 1 % of pinned — the copies are LINK-limited, not staging-limited, so the usual
+  the copies are LINK-limited, so the usual "pageable costs 2×" rule does not apply; the x4 lives in
-  "pageable costs 2×" rule does not apply. The card and the port beneath it are both wider than x4,
+  the adapter or cable (card and port are both wider). For any engine that STREAMS weights across
-  so the x4 lives in the adapter or cable, and the link idles at 2.5 GT/s until load. Consequence
+  it: a 3 MiB slab costs 474 µs, so a 111 GiB A3B model needing ~1.70 GiB of active experts per
-  for any engine that STREAMS weights across it: a 3 MiB slab costs 474 µs, so a 111 GiB A3B model
+  token lands at 3.7 tok/s against 20.46 tok/s with the experts COMPUTED in host RAM
-  needing ~1.70 GiB of active experts per token lands at 3.7 tok/s, against 20.46 tok/s for the same
+  (`-ot exps=CPU`, where the link carries activations only). Streaming loses; RAM does not care.
-  model with the experts COMPUTED in host RAM (`-ot exps=CPU`, where the link carries activations
+  [[x4-pair-experiment]]
-  only). Streaming over PCIe loses; computing in RAM does not care. [[x4-pair-experiment]]
 
 
 **ROCm / GPU**
@@ -588,12 +584,12 @@ linked runbooks.
   can boot before a human reaches it. [[frank-archon]]
 - **Silent RAM corruption outranks every software theory.** frank-archon's long-conversation
   repetition/garbage was marginal DDR5 at 4800 MT/s on four DIMMs, fixed by dropping to 4600.
-  Consumer non-ECC memory raises NO MCE for a silent flip, and a short in-band pattern check
+  Consumer non-ECC raises NO MCE for a silent flip, and a short in-band pattern check (60 s over
-  (60 s over 8 GiB of 246 GiB) is not a stability proof — combined, those two instruments said
+  8 GiB of 246 GiB) is not a stability proof — both instruments said "ruled out" while the RAM was
-  "ruled out" while the RAM was the cause, and "4800 is the JEDEC floor" is a rating, not a margin
+  the cause, and "4800 is the JEDEC floor" is a rating, not a margin guarantee. When the fault is
-  guarantee. When the fault is data-shaped (NaN states, garbage tokens) and a software revert does
+  data-shaped (NaN states, garbage tokens) and a software revert does not move it, price in the RAM
-  not move it, price in the RAM before the runtime — and remember a revert tested under the same
+  before the runtime; a revert tested under the same marginal hardware is itself confounded.
-  marginal hardware is itself confounded. [[log]]
+  [[frank-archon]]
 
 **Runtime-specific guards**
 
@@ -724,11 +720,10 @@ linked runbooks.
   `CMakeCache.txt` — never from the build flags you remember passing. [[frank-zealot]]
 - **A llama.cpp endpoint is invisible to Prometheus unless `--metrics` is passed**: without it
   the engine answers **501 Not Implemented** on `/metrics`, so the scrape fails while `/health`
-  answers 200 and `LlmEndpointDown` fires beside a healthy service — that is how `frank-zealot`
+  answers 200 and `LlmEndpointDown` fires beside a healthy service. halogen serves `/metrics`
-  sat red across four payload generations. halogen serves `/metrics` natively, so the asymmetry
+  natively, so the asymmetry reads as "only the 3090 is broken". Diagnose with `lastError` from
-  reads as "only the 3090 is broken". Diagnose with `lastError` from `/api/v1/targets`, not the
+  `/api/v1/targets`, not the alert list, and note Alertmanager has no persistence: a reboot
-  alert list, and note Alertmanager has no persistence: a reboot re-stamps `startsAt` on an alert
+  re-stamps `startsAt` on an alert that has fired for weeks. [[Observability]] [[frank-zealot]]
-  that has fired for weeks. [[Observability]]
 - **A leftover bench drop-in silently hijacks the next deploy, and no health gate sees it.** The
   sanctioned bench-window method (a transient `Environment=FRANK_<SVC>_ENV_FILE=/tmp/<arm>.env`
   drop-in sourcing the repo env) outlives the bench unless retired: a leftover