TokenPak v1.3.10

Changed — Phase A of MVP Gap Closure (Working Product Readiness initiative)

Ratified by the maintainer 2026-04-24. Closes three P0 gaps and establishes honest path-specific savings framing before any launch work opens. Details at internal docs.

P0-5 — path-specific savings framing (this is the big one): The fresh-install product audit surfaced that tokenpak on Claude Code traffic measures ~2% savings even with compression working correctly — because Anthropic's server-side cache already captures 95% of the reducible token pool. The ≥30% CI-pinned floor on the agent-style fixture is honest; the aggregate ~2% on cache-heavy real workloads is also honest; users evaluating on Claude Code alone concluded the claim was false.

Per the maintainer's canonical ruling: do NOT flatten into a single 30–50% claim. Use path-specific language.

• README.md — headline rewritten from "Cut your LLM token spend by 30–50%" to "Up to 90%+ savings on direct API/CLI and other favorable uncached workloads". New "How we report savings" section explains the two paths explicitly: direct API / CLI / uncached workloads land in the 90%+ band; provider-cached flows show lower incremental gains because the cache already did the heavy lifting.
• "What's included" — "50 built-in compression recipes" claim removed. Replaced with an honest description: deterministic compression pipeline + route-class policy presets + custom recipe authoring via tokenpak recipe create/validate/test/benchmark.
• Reproduction footer — reworded to "Reproduce the savings floor locally: make benchmark-headline (asserts ≥30% reduction on a pinned agent-style fixture; the CI-enforced floor, not the ceiling)." The benchmark is honest; the prior phrasing implied a ceiling that wasn't the point.

P0-1 — recipes reality alignment: The recipes_oss/*.yaml catalog referenced in pyproject.toml package_data did not ship with v1.3.9 (verified absent from the installed wheel). Rather than resurrect an archived recipe pack, the shipped architecture is treated as source of truth.

• pyproject.toml — removed recipes_oss/*.yaml from package_data. Added services/policy_service/presets/*.yaml + agent/compression/slot_definitions.yaml which actually ship.
• tokenpak demo --list — no longer tells the user to reinstall (which wouldn't fix anything). New message explains: "TokenPak's compression is a deterministic pipeline, not a bag of YAML recipes," and points at tokenpak demo (live compression), tokenpak recipe --help (custom recipes), and tokenpak status (see it running).
• README — "50 recipes" claim dropped; services/policy_service/presets/*.yaml coverage (9 route-class presets) documented honestly.

P0-2 — setup command in quick help: tokenpak --help (first-run path) previously showed [start, demo, cost, status] — setup was registered in _COMMAND_GROUPS (A1, v1.3.8) but not in _QUICK_COMMANDS, so a first-time user seeing the quick help had no hint to run it. README's "One command to configure" promise was structurally invisible.

• tokenpak/cli/_impl.py — prepended "setup" to _QUICK_COMMANDS so the first command a new user sees is the one the README promises.

Tests: no new tests added (the _QUICK_COMMANDS change is covered by the existing test_setup_registration.py smoke; the other two are packaging + copy). Conformance matrix 102/102 green; full suite green.