chore(deps): bump tiktoken from 0.12.0 to 0.13.0

[dependabot]

Bumps tiktoken from 0.12.0 to 0.13.0.

Changelog

Sourced from tiktoken's changelog.

[v0.13.0]

• Update fancy-regex for significantly increased performance
• Branch byte pair encoding to fix performance on unusual input
• Fix AttributeError caused by incomplete redaction of experimental code
• Update version of pyo3
• Update version of optional dependency blobfile

Commits

• fa8b65d Partial sync of codebase (#543)
• dcb3928 [codex] Pin GitHub Actions workflow references (#515)
• 6ec8149 Partial sync of codebase (#495)
• 6ad70b2 Upgrade GitHub Actions for Node 24 compatibility (#472)
• See full diff in compare view

[github-actions]

NeuralMind self-benchmark

Status: PASS — floor 4×, measured 6.6×.

Phase 1 — Reduction on committed fixture

• Average reduction: 6.6×
• Top-k retrieval hit rate: 71.7%
• Naive baseline: 47,360 tokens (all fixture files concatenated)
• NeuralMind total: 7,296 tokens across 10 queries
• Estimated monthly savings @ 100 queries/day on Claude 3.5 Sonnet: ~$36.06

#	Query	Shape	Naive	NeuralMind	Ratio	Hit
1	auth-flow	cross-file	4,736	726	6.5×	33.3%
2	api-endpoints	focused	4,736	714	6.6×	100.0%
3	billing-flow	cross-file	4,736	710	6.7×	33.3%
4	user-storage	cross-file	4,736	825	5.7×	50.0%
5	jwt-verify	focused	4,736	590	8.0×	100.0%
6	stripe-webhook	focused	4,736	599	7.9×	100.0%
7	create-user	cross-file	4,736	855	5.5×	50.0%
8	refund	focused	4,736	732	6.5×	100.0%
9	db-choice	identity	4,736	834	5.7×	100.0%
10	invoice-send	cross-file	4,736	711	6.7×	50.0%

Phase 2 — Learning uplift

• Memory events logged: 20
• Learned patterns: 14
• Reduction ratio after neuralmind learn: 6.5× (Δ -0.09× vs. cold)
• Top-k hit rate after learning: 75.0% (Δ +3.3 points vs. cold)

Note: uplift numbers on a 500-line fixture are intentionally modest — the point is to
verify the learning mechanism persists and applies. On real production repos the lift
is larger; this test only catches regressions in persistence.

Phase 3 — Synapse recall A/B (same warm graph, recall off vs on)

• Synapse edges after seeding co-editing sessions: 3610
• Top-k hit rate: 71.7% off → 83.3% on (Δ +11.7 points)
• Reduction ratio: 6.6× off → 6.6× on (Δ -0.07× — budget-neutral by design)

This isolates the Hebbian synapse layer from the learned_patterns reranker in
Phase 2. The hit-rate delta shows associative recall surfacing co-edited modules a
purely textual search ranks lower; the near-zero reduction delta confirms it does so
without spending extra tokens (recalled nodes displace the weakest hits, not add to them).

Assumptions

• Baseline: every .py file in tests/fixtures/sample_project/ concatenated.
• Tokenizer: tiktoken GPT-4o encoding (per-model breakdown in multi_model.json if generated).
• Pricing: Claude 3.5 Sonnet input @ $3.0/MTok.
• Regression floor: 4× — well below NeuralMind's typical 40–70× on real repos.

Per-model token reduction

Model	Tokenizer	Naive	NeuralMind	Ratio	Source
GPT-4o / GPT-4o-mini	tiktoken o200k_base	4,739	779	6.1×	measured
GPT-4 / GPT-3.5-turbo	tiktoken cl100k_base	4,710	770	6.1×	measured
Claude 3.5 Sonnet	estimated: GPT-4o × 1.08 — install anthropic for an exact count	5,118	841	6.1×	estimated
Llama 3 (70B)	estimated: GPT-4o × 1.22 — Llama tokenizer requires model weights; estimate based on published vocab ratios	5,781	950	6.1×	estimated

Rows marked measured use the provider's real tokenizer. Rows marked
estimated apply a published vocab-size correction to the GPT-4o count —
honest approximations, not hardcoded claims.

---

Automated by .github/workflows/ci-benchmark.yml — regenerate locally with python -m tests.benchmark.run.

[dfrostar]

@dependabot rebase

---

Generated by Claude Code