v0.1.40

fix: shim-shadowing field bug + from-scratch synthesis upgrades (v0.1…

v0.1.39

feat: optimize_build (holistic optimizer) + craft_item (crafting syst…

v0.1.38

fix: optimize_passives full-tree default + unique base parse (v0.1.38)

Dogfooding v0.1.37 live verified all of its new capabilities work (crit terminology fix,
support-pool fix, ascendancy auto-allocation + reporting, plan_gear auto-base, relevant_uniques,
the trigger guardrail). It surfaced two fixes:

- optimize_passives: the MCP tool defaulted points=3, so a bare optimize_passives() allocated only
  3 passive points and then reported pointsRemaining:0 — which reads as 'tree fully allocated', so a
  model building a build would ship a 3-point tree. The default is now 0 (the whole tree, the usual
  intent; pass a positive number only to cap it), and the shim reports pointsRemaining as the
  build's TRUE unspent passive points so a capped call can never misread as done.

- pipeline: a unique's 'Source:' drop line was mistaken for its base (Hand of Wisdom and Action read
  'Source: Drops from...' instead of 'Furtive Wraps'). UNIQUE_META_RE now filters Source: lines
  (1 of 433 uniques affected; corrects on the next corpus rebuild).

Full gate green (ruff, mypy, pytest).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

v0.1.37

feat: pinnacle-build improvements + dogfood fixes (v0.1.37)

From-scratch builds now produce a complete, resist-capped, engine-verified set; several
diagnostic/correctness bugs surfaced by dogfooding are fixed.

Correctness:
- optimize_supports: measurement-based candidate pool (penetration etc. no longer excluded
  by tag-count), recovering ~38% DPS on the dogfood build
- solver: PoE1 'Critical Strike Chance' to PoE2 'Critical Hit Chance' (the engine silently
  ignored the stale term, hiding the crit lever in rank_levers/solve_for) + a standing
  terminology-audit test
- shim: ascendancy is a SEPARATE 8-point pool. optimize_passives auto-allocates it without
  stranding passive points or over-allocating; get_build reports ascendancyPoints + warns
- meta-trigger guardrail: flag engineLimitation when a build uses an energy-meta trigger gem
  (Cast on Critical etc.) the pinned engine cannot model, instead of faking the number

Features:
- plan_gear: auto-base empty armour/jewellery slots (a whole set from one weapon) + min_ehp floor
- relevant_uniques: build-aware unique + unique-jewel discovery (tools 61 to 62)

Knowledge: archetype-commit guidance, what-the-engine-can-and-cannot-model, and unique-awareness
woven into BUILD_ADVICE and the build_from_goal prompt.

Full gate green (ruff, mypy, pytest 130).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

v0.1.36

v0.1.36: instruction + top-level docs refreshed for the new tools

v0.1.35

v0.1.35: plan_gear cross-slot budget allocation

v0.1.34

v0.1.34: craftable jewels + support-set optimizer

v0.1.33

v0.1.33: gear-craft attainability + craft-effort

v0.1.32

v0.1.32: blended gear crafting (goals) + rank_upgrades

v0.1.31

v0.1.31: optimize_item fixes + sharper build instructions