Deeper MahaMantra substrate integration — steward cognition layer

[kimeisele]

The Vision

steward-protocol/vibe_core has massive infrastructure that steward barely uses:

Used

• MahaAttention (O(1) routing) ✓
• MahaCompression (seed extraction) ✓ (in ExploreCapability)
• HebbianSynaptic (learning) ✓
• SiksastakamSynth (booted but passive)
• AntarangaRegistry (512-slot O(1) state — booted but passive)

Not used

• MahaBuddhi (deterministic thinking) — steward delegates everything to LLM
• CognitiveWeaver (State↔Knowledge bridge) — never called
• MahaComposition (5-scorer ranking) — only in ExploreCapability
• OpCode system (16 operations) — not mapped to steward actions
• Guna classification at action level — only at file level
• InferProtocol (atomic inference) — never used
• CognitiveKernel (rate-limited thinking) — never used

Impact

These are not "nice-to-haves." They are the SUBSTRATE that makes steward
a governed intelligence instead of a script runner. Without them, steward
is just a glorified cron job with a healer attached.

The Siksastakam 7 effects are literally the computing principles:

1. Cache invalidation (cleanse heart mirror)
2. Zero entropy routing (extinguish forest fire)
3. Graceful degradation (spread moonlight)
4. Live data structures (life of knowledge)
5. Infinite scalability (expand bliss ocean)
6. Atomic transactions (full nectar each step)
7. Total transformation (bathe entire self)

Steward should embody ALL 7, not just #1.

[kimeisele]

WCFA Audit Result

25/31 services ACTIVE, 6 WCFA'd (booted but 0 call sites):

Service	Purpose	Impact
SVC_DIAMOND	TDD RED/GREEN gates	HIGH — tests aren't gated
SVC_INTEGRITY	Boot-time validation	MEDIUM — integrity checked once at boot, not ongoing
SVC_MAHA_LLM	Deterministic semantic engine	HIGH — steward uses LLM instead of deterministic semantics
SVC_OUROBOROS	Self-healing orchestrator	HIGH — immune system partially replaces this
SVC_PROMPT_CONTEXT	Dynamic system prompt	MEDIUM — system prompt is static
SVC_SIKSASTAKAM	7-beat cache lifecycle	LOW — cache cleansing wired to immune (#16)

Priority activation order:

1. SVC_MAHA_LLM → deterministic intent classification before LLM
2. SVC_PROMPT_CONTEXT → dynamic system prompt from live state
3. SVC_DIAMOND → TDD gates in the healer pipeline
4. SVC_OUROBOROS → merge with StewardImmune (overlapping responsibility)

Also: CognitiveWeaver is NOT booted at all — it's only in vibe_core.

[kimeisele]

SVC_MAHA_LLM Activation Analysis

services.py line 296: # 18. MahaLLMKernel — DEFERRED (registered but not consumed by agent loop)

This was INTENTIONALLY deferred, not accidentally WCFA'd. The MahaLLM (O(4) holographic router) needs to be wired into the AgentLoop BEFORE the LLM call — as a deterministic pre-filter:

User intent → MahaLLM.route() → if deterministic match → execute directly (0 tokens)
                               → if no match → fall through to LLM (normal path)

This is the CORE of "architecture > LLM" — the North Star.

Where to wire: steward/loop/engine.py — before _call_llm(), check MahaLLM first.
What to register: All intent handlers, tool names, common patterns.
Impact: Every deterministic intent saves a full LLM roundtrip.

This is the single highest-impact activation for #17.