[Draft] Integrate ImpactIA methodology

[benoit-cty]

What Claiude Fable 5.1 did

The ImpactIA work is now integrated on the feat/impactia branch (uncommitted, 27 files). Lint passes, the docs build, and the suite is at 211 passed. The 4 failures in tests/test_openai.py (Responses API) fail identically on the untouched branch: the recorded cassette lacks cache_write_tokens, which the current OpenAI SDK requires, so they are unrelated.

Core inference model, on by default (llm.py):

• Input tokens now matter. Pre-fill latency is modeled from the input token count (Agrawal et al. coefficient, scaled by each model's OpenRouter TTFT, as ImpactIA proposes). All seven tracers pass the provider-reported input tokens.
• Datacenter network equipment (firewall, router, switch): energy plus embodied GWP, allocated like the server.
• Datacenter building embodied GWP factor per kWh of IT energy.
• ADPe and PE are not extended for these two items because ImpactIA only provides GWP data; the docs say so.

New experimental training phase (training.py): final run, R&D ×4, dataset storage, allocated per output token over the model's lifetime tokens. It requires a model release date (added to 146 models in models.json) and a provider compute capacity (Anthropic, Google, Mistral, OpenAI). Otherwise a new training-not-modeled warning is emitted. It is reported separately and not added to the totals, so existing numbers only shift by the network and building additions.

Also fixed a real bug on the way: RangeValue.__rtruediv__ computed the wrong interval.

Docs: new llm_training.md, updated llm_inference.md, tutorial, warnings page, nav, and a blog post whose byline is a placeholder (samuelrince).

Something to report back to the ImpactIA team

Their spreadsheet has a unit inconsistency in cell Calcul!N115: provider compute capacity is stored in kW but multiplied by FLOPS per W without the ×1000. This divides the estimated lifetime token volume by 1000 and is what produces their headline "training = 13% median". I implemented the guide's method with consistent units, so the share EcoLogits reports is much smaller:

Model | Training share of GWP | Energy change, 100k input tokens -- | -- | -- gpt-4.1-mini | 0.19% | +5.9% claude-opus-4-6 | 0.04% | +15.2% mistral-large-latest | 7.4% | +4.0% gemini-2.5-flash | 0.01% | +10.3%

The token-volume estimate remains the weak point of the method in both directions, which the methodology page states plainly. All parameters can be overridden, including model_lifetime_output_token_count.

Deliberately left out

• RAG embedding stage: worth doing as an embeddings methodology, but needs embedding models in the repository and tracer support for embedding endpoints.
• "Front applicatif": SNCF-specific hosting assumption, out of scope for an API-level library.
• AWARE-weighted embodied water: a different indicator from WCF in liters, so I did not merge it.
• ResilioDB embodied factors: kept EcoLogits' sourced values.

Points needing your decision

• License: ImpactIA's guide and spreadsheet are CC BY-NC-SA. I re-implemented the formulas and wrote original documentation text with attribution, but you may want a legal check before merging into an MPL-2.0 project.
• Release dates are missing for a few models (claude-opus-4-7, some Gemini 3.1 previews, rolling -latest aliases), so they get the warning instead of a training estimate.
• Nothing is committed; review the diff and commit when satisfied.

[benoit-cty]

ccusage measure:
| Agent │ Models │ Input │ Output │ Cache Create │ Cache Read │ Total Tokens │ Cost (USD) │
| - Claude │ - fable-5-1 │ 996 │ 144,004 │ 430,255 │ 8,806,508 │ 9,381,763 │ $18.02 │

[samuelrince]

Hey @benoit-cty, thanks I think I'd rather meet them and discuss some parts of the methodology before reviewing this PR!

[XavierVerne]

Thanks for this detailed pull request and all the work behind. A first few things on what you explained :

1. Please let us know if there is a blocking condition regarding our licence, so as for us all to find a way. The purpose was to facilitate external integration while preventing people to close further improvments.
2. thanks for the bug, fixing in progress.
3. I understood that Samuel and Rosendo and Thibaud are about to talk to each other. May be you could join in ? refer back to Samuel I don't have all infos. This could accelerate things.