Release 0.11.2

Token Terminator v0.11.2 — Sol counting and reliable TT discovery

This repair addresses two blockers left in 0.11.1. It does not change JEV keys, semantic policies, the evidence vault, Context IR, or the main-call-only boundary.

GPT-6.1-Sol

The exact gpt-6.1-sol ID and known OpenAI/Codex/OpenRouter-qualified forms now resolve to o200k_base when the installed tiktoken model-name table lacks the ID. An upstream mapping still wins, and an explicit operator encoding remains supported. Unrelated providers, invented versions and sibling names are not inferred.

Five synthetic inputs were checked against OpenAI's documented input-token counter: English, multilingual/emoji text, code/numbers, structured JSON and literal special markers. All matched local content counts plus the same six-token message framing. The recorded corpus distinguishes o200k_base from cl100k_base. This is empirical compatibility evidence, not an upstream tiktoken declaration or proof about every possible input. TT still measures complete canonical request JSON, not hidden provider framing or billed tokens. No generation requests were used in that check.

Concurrent Hermes discovery

The affected shared loader can return a module before its initialization completes, and its slug-only cache can reuse one profile's module for another. TT's explicit install-context-engine command now applies a source-hash-guarded repair to the reviewed affected loader. TT loads are synchronized, recursive discovery cannot consume a partial module, and each resolved profile path has a distinct namespace. Other plugins use the original loader. Failed initialization can be retried.

This is a narrow host-file compatibility patch, not merely a plugin import change: the race occurs before the plugin can run. An adjacent byte-exact backup is kept. Unknown or modified host implementations are refused rather than overwritten. The Python installer API and install-dashboard do not patch the host by default. No changes occur on package import, and configuration/history are not rewritten.

Install

Use the interpreter/environment actually loaded by the managed Hermes launcher:

python -m pip install --upgrade 'git+https://github.com/AronAxe/Token-Terminator.git@v0.11.2'
python -m rtk_hermes_plus.cli install-context-engine

The second command reports discovery_repair; for an affected installed host it must say repaired or already_repaired. Then restart the appropriate idle backend through its supervisor once. A package download alone cannot repair a running process. Existing explicit engine selection and JEV/IR opt-ins remain unchanged.

Regression checks exercise the actual GPT-6.1-Sol model name without encoding overrides, JEV selection, strict complete-request reduction and exact recovery. Discovery checks include controlled overlapping initialization, 64 simultaneous loads, profile separation, recursive loading, failed-load retry and safe installation. A dedicated CI job runs those same tests against the actual pinned Hermes loader source, not an optional skipped module. Full CI results are recorded on the PR.

The Rust companion is version-aligned only. Existing tags, Terminator artwork and schematics remain unchanged. See docs/TOKENIZER_COMPAT.md and docs/HERMES_DISCOVERY.md.