Add Noosphere Engine foundation (Phase 0 + Phase 1)

[user1303836]

Summary

Implements the foundational infrastructure for the Noosphere Engine feature, covering Phase 0 (core package structure) and Phase 1 (metrics computation layer).

Phase 0 - Core Infrastructure:

• SQLAlchemy models (10 tables): NoosphereGuildState, MessageEmbedding, EgregoreSnapshot, UserBioelectricState, SoundscapeSnapshot, AttractorSnapshot, ArchiveEntry, ArchiveLink, CrystalRoom, GuildMetricsBaseline
• Shared dataclasses: ProcessedMessage, CommunityStateVector
• Constants and enums: PHI, Fibonacci sequence, ComputationMode (10 modes), PathologyType (10 types), MessageClassification
• EmbeddingService: async wrapper around sentence-transformers (paraphrase-multilingual-MiniLM-L12-v2, 384-dim)
• OutputGovernor: O(1) sidechain gain computation with token bucket and cooldown
• SoundscapeMonitor: anthrophony/biophony/geophony message classification

Phase 1 - Metrics and Computation:

• MetricsComputer: three-tier scheduling (hourly/daily/weekly) producing CommunityStateVector
• WelfordAccumulator: online mean/variance with z-score normalization and sigmoid mapping
• Egregore Index: weighted composite of coherence (0.4) + convergence (0.3) + concentration (0.3)
• Archive decay: stigmergic storage with reference-extended half-life (base 168h, 1.5x per reference)
• FibonacciScheduler: golden-angle jitter for quasiperiodic scheduling
• PhiParameter: golden ratio mode weight oscillator

Dependencies added: sentence-transformers, chromadb, numpy, scipy, networkx

Test plan

• 90 unit tests covering all modules (all passing)
• ruff check passes
• ruff format passes
• mypy passes with no issues (14 source files checked)
• SQLAlchemy models tested with in-memory SQLite
• Numerical algorithms verified (Welford convergence, coherence bounds, entropy normalization)

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.

[user1303836]

Code Review: PR #187 -- Noosphere Engine Foundation (Phase 0 + Phase 1)

Verdict: Changes Requested -- 7 issues to address before merge, 5 positive observations.

---

Issues Requiring Changes

1. Enums use enum.Enum instead of str, Enum -- constants.py

Per our alignment agreement, all enums should use str, Enum as base classes for JSON serialization compatibility. Currently ComputationMode, PathologyType, and MessageClassification all inherit from bare enum.Enum.

# Current
class MessageClassification(enum.Enum):
    ANTHROPHONY = "anthrophony"

# Expected
class MessageClassification(str, enum.Enum):
    ANTHROPHONY = "anthrophony"

2. CommunityStateVector missing agreed fields -- shared/data_models.py

The dataclass is missing fields that dev-analytics and dev-features depend on:

• sentiment_alignment: float = math.nan
• interaction_modularity: float = math.nan

These were agreed upon during interface alignment. Use math.nan defaults so Phase 1 code doesn't need to compute them yet, but downstream consumers can check for them.

3. datetime.utcnow() is deprecated -- metrics_computer.py:58

datetime.utcnow() is deprecated since Python 3.12. The existing codebase uses datetime.now(UTC) (see database/models.py). Change to match:

from datetime import UTC, datetime
# ...
now = datetime.now(UTC)

4. TYPE_CHECKING imports used for runtime types -- shared/data_models.py

numpy and MessageClassification are imported under TYPE_CHECKING, but they're used at runtime in dataclass fields. This works today because of from __future__ import annotations, but it's fragile -- any code that inspects field types at runtime (dataclasses.fields, serialization libraries, isinstance checks) will break. Move these to unconditional imports.

5. chromadb listed as dependency but unused -- pyproject.toml

chromadb appears in the dependency list but no code imports it. If it's planned for a future phase, remove it now and add it when actually needed. It's a heavy dependency.

6. Duplicate [dependency-groups] section -- pyproject.toml

There are two [dependency-groups] sections in pyproject.toml. TOML spec says duplicate tables are invalid. Some parsers silently merge them, others may error. Consolidate into one section.

7. MetricsComputer unbounded memory growth -- metrics_computer.py

_messages and _embeddings dicts append indefinitely with no pruning. For a long-running bot, this will eventually consume all available memory. Add a rolling window or clear after each computation cycle:

def compute_hourly(self, guild_id: int) -> CommunityStateVector:
    messages = self._messages.pop(guild_id, [])
    embeddings = self._embeddings.pop(guild_id, [])
    # ... compute ...

Or cap at a configurable max size (e.g., last 1000 messages per guild).

---

Positive Observations

• OutputGovernor (shared/output_governor.py): Clean O(1) gain computation with token bucket pattern. Per-channel tracking, gain floor of 0.1, stochastic send decision -- all match the self-quieting design principle well.

• WelfordAccumulator (shared/baseline.py): Numerically stable online mean/variance. Correct use of population variance (N not N-1 for normalization purposes). Sigmoid clamped at +/-10 to avoid overflow.

• EgregoreIndex default (shared/egregore.py): Returns 0.5 when baseline count < 2, avoiding division-by-zero and giving a neutral starting point. Weighted composite (0.4/0.3/0.3) matches spec.

• Archive decay math (shared/archive_decay.py): effective_half_life = base * (1.5^references) correctly extends relevance for frequently-referenced entries. Fidelity floor of 0.01 prevents complete erasure. Interaction boost with exponential decay is well-designed.

• FibonacciScheduler (shared/fibonacci.py): Golden-angle jitter with max(0.5, ...) floor prevents degenerate intervals. Clean cycle-through-sequence design.

---

@greptile

[greptile-apps]

Thanks for the detailed review! All 7 issues are clear and actionable. I'll address them before requesting re-review:

1. ✅ Change enums to str, Enum base classes
2. ✅ Add sentiment_alignment and interaction_modularity fields with math.nan defaults
3. ✅ Replace datetime.utcnow() with datetime.now(UTC)
4. ✅ Move numpy and MessageClassification out of TYPE_CHECKING block
5. ✅ Remove chromadb dependency
6. ✅ Consolidate duplicate [dependency-groups] sections
7. ✅ Fix memory leak by adding .pop() to clear messages/embeddings after computation

Appreciate the positive feedback on the implementation patterns! Will push fixes shortly.

[user1303836]

Re-review: All 7 Items Addressed

I've verified every fix against the original review. All items are resolved.

Original Issues -- Status

#	Issue	Status
1	Enums use bare enum.Enum	Fixed -- ComputationMode, PathologyType, MessageClassification all use (str, enum.Enum) base (constants.py lines 24, 37, 50)
2	CommunityStateVector missing Phase 4 fields	Fixed -- sentiment_alignment, interaction_modularity, fractal_dimension, lyapunov_exponent, gromov_curvature added with math.nan defaults (data_models.py lines 42-46)
3	datetime.utcnow() deprecated	Fixed -- replaced with datetime.now(UTC) in metrics_computer.py (line 68). No remaining utcnow references in source or tests.
4	TYPE_CHECKING imports for runtime types	Fixed -- datetime, numpy, MessageClassification now imported at runtime with # noqa: TC001/TC002/TC003 comments (data_models.py lines 5, 7, 9)
5	Unused chromadb dependency	Fixed -- removed from both [project.dependencies] and [[tool.mypy.overrides]]. Zero references to chromadb in the codebase.
6	Duplicate [dependency-groups] section	Fixed -- zero occurrences of [dependency-groups] in pyproject.toml
7	MetricsComputer unbounded memory	Fixed -- MAX_MESSAGES_PER_GUILD = 5000 and MAX_EMBEDDINGS_PER_GUILD = 5000 constants added (metrics_computer.py lines 18-19). ingest_message() trims both lists when exceeded (lines 42-49). Two new tests verify bounds.

Verification Checklist

• data_models.py imports are now runtime (not TYPE_CHECKING), preventing fragile annotation-only behavior
• All IDs remain str throughout
• pyproject.toml is clean: no duplicate sections, no phantom dependencies, mypy overrides list is accurate
• 94 noosphere tests (including 2 new bounds tests)

Minor Note (non-blocking)

Foundation uses (str, enum.Enum) tuple base class while analytics (PR #188) uses enum.StrEnum. Both serialize identically, but if you want consistency across branches, pick one style. StrEnum is the more modern Python 3.11+ approach. Not blocking since both work correctly.

Verdict: Approve

This PR is clean and ready to merge.