docs: RULE #1 — no automated fallbacks, fail fast/loud (CLAUDE.md + AGENTS.md)

Every op goes through ONE clear path; when it does not work as expected/errors/misbehaves,
RAISE IMMEDIATELY with where+what and crash so the bug surfaces and we fix the root cause.
Forbidden: silent/automated fallbacks (try/except->degraded path, except:pass, eager-replaces-
fused, gated-off-by-default, clamp-instead-of-raise, return zeros/garbage, best-effort). Fail-
closed by raising a clear error is correct; silent fallback producing degraded/wrong/zero output
is forbidden. Also fixed AGENTS.md tensor rule (was permitting 'fall back to production path').

Author: jeweldave

AGENTS.md

@@ -1,5 +1,15 @@
 # AGENTS.md — cppmega.mlx agent contract
 
+## RULE #1 (HIGHEST PRIORITY) — NO AUTOMATED FALLBACKS. FAIL FAST, FAIL LOUD.
+
+Every op goes through ONE clear path. When the clear path does not work as expected, errors,
+or misbehaves, **RAISE IMMEDIATELY** with WHERE + WHAT failed and crash — surface the bug, fix
+the root cause. FORBIDDEN: silent/automated fallbacks (try/except→degraded path, except:pass,
+eager-replaces-fused, MSL-instead-of-tvm-ffi, gated-off-by-default, clamp-instead-of-raise,
+return zeros/garbage, best-effort, silent precision/shape downgrade). Fail-closed by RAISING a
+clear error is correct; a silent fallback producing degraded/wrong/zero output is forbidden.
+See CLAUDE.md. This overrides every rule below.
+
 ## Tensor memory rule
 
 - Wrappers/adapters must not silently allocate or copy large tensors.
@@ -8,8 +18,8 @@
   design as wrong and keep looking.
 - Prefer passing references to existing GPU buffers, fusion, views/broadcasts,
   or IR-level lowering through the TileLang/TVM/tvm-ffi pipeline.
-- If zero-copy/fused lowering is not currently possible, fail explicitly or
-  fall back to the existing production path without pretending Path C handled it.
+- If zero-copy/fused lowering is not currently possible, FAIL EXPLICITLY (raise with
+  where + what). Do NOT silently fall back to another path (Rule #1).
 - Dtype/shape mismatches must be solved at the right level:
   - first look higher in the graph and make the producer create the tensor in
     the required dtype/shape from the start;

CLAUDE.md

@@ -0,0 +1,25 @@
+# CLAUDE.md — cppmega.mlx
+
+## RULE #1 (HIGHEST PRIORITY — overrides everything else) — NO AUTOMATED FALLBACKS. FAIL FAST, FAIL LOUD.
+
+Every operation goes through ONE clear path. **When the clear path does not work as expected,
+errors, or misbehaves, the code MUST RAISE IMMEDIATELY** with a precise message saying **WHERE
+and WHAT** failed, and crash — so the bug surfaces and we fix the ROOT CAUSE.
+
+**FORBIDDEN** — automated/silent fallbacks that hide a failure:
+- `try/except` → degraded/alternate path; `except: pass`; "best-effort"; silent slow-path.
+- eager-replaces-fused, MSL-adapter-instead-of-tvm-ffi, gated-off-by-default, watchdog→fallback.
+- clamp-instead-of-raise; return zeros/garbage; silent shape/dtype/precision downgrade.
+- any code path that papers over a broken clear path instead of surfacing it.
+
+**REQUIRED on failure:** `raise` with where + what → then we debug and fix the bug.
+Fail-closed by **raising with a clear error IS correct** ("падать с сообщением"). A SILENT
+fallback that produces a degraded/wrong/zero result is what is forbidden.
+
+This is the first and main rule. Existing automated fallbacks must be found and removed; route
+everything through the single clear path that fails loud.
+
+## Tensor memory rule (see AGENTS.md)
+- Wrappers/adapters must not silently allocate, cast, or copy large tensors to satisfy a
+  boundary. If a design seems to require it, the design is wrong — keep looking (this is a
+  corollary of Rule #1: no silent papering-over).