0.3.0 — Tactical mode port (engine, mode input on the tools, tactical goldens)

The tactical-mode port: the engine now covers both of upstream's engagement
plans. Orbit-mode behavior is unchanged — the orbit fixtures regenerate
byte-identically (their _meta.source_commit pin advances to upstream
843a2c5, the commit that introduced tactical mode without touching orbit
behavior).

Added

• Tactical mode in the engine (src/engine/tactical.ts plus a
  mode: "orbit" | "tactical" argument on Simulation / runEngagement,
  mirroring the browser's resetSim(seed, mode)): the multi-FPV sortie
  stream — per-seed launch plans and aim points, per-airframe EMCON keying,
  the pilot-station launch scheduler with commit holds, the strike/hunter
  FSM sharing the orbit attack run, and the sim's own STALEMATE end state
  (outcome reason packages_expended). Both modes share every RNG draw
  through emplacement; tactical results add the objective, per-airframe
  package state, sortie tallies and the killer id.
• mode input on the tools: run_engagement, sweep_seeds and
  compare_configs accept mode ("orbit" default); TACTICAL.* overrides
  (package size, pilot stations, reserve-or-retask, launch spacing,
  objective geometry — including the boolean RESERVE_HUNTER) join the
  validated parameter table and get_config_schema; describe_model gains
  a tactical_mode section; sweeps over tactical results add per-side
  strikes_delivered distributions.
• Second golden-fixture set (test/fixtures/golden-seeds-tactical.json,
  the six featured tactical seeds) with matching golden-master tests, and
  tactical coverage in the fixture generator. Parity for the port was
  additionally cross-checked against the vm-run original browser sim over
  seeds 1–500: outcomes, event logs, fix floats and per-airframe positions
  all match exactly, reproducing the documented 27% / 15% / 58% split.

Changed

• CI: the upstream-drift gate now regenerates and compares BOTH fixture
  sets, so a tactical-behavior change upstream (or an engine regression in
  either mode) is caught the same way orbit drift always was.

Docs

• README, DESIGN_NOTES.md and the pinned docs/upstream/DESIGN_NOTES.md
  copy re-synced for the two-mode era (SNAPSHOT.md pin: 843a2c5).

0.2.3 — Per-PR parity gate (upstream-drift on every PR, required on main)

No engine or tool behavior changes; golden fixtures unchanged. CI and
docs only: the upstream-drift parity check now runs on every pull request
and is a required status check on main (paired with fpv-sim's new
parity workflow, so the two repos cannot diverge at merge time), and
describe_model / the pinned DESIGN_NOTES copy are corrected against the
code.

Docs

• describe_model text (src/server/model.ts) corrected to match the
  code, in step with upstream fpv-sim's DESIGN_NOTES.md fix (fpv-sim
  PR #20): pathAtten samples 13 interior points and tops out near 3.9
  (the cap of 6 is never reached); WLS weights floor range at 300 m; the
  geometry-penalty cut angle is between the two strongest sensors'
  bearings to the current estimate, not their mean LOBs; the terminal
  search is an outward spiral, not an "expanding search" of unspecified
  shape; and the determinism entry names every main-stream consumer
  (including the display-only enemy-drone track noise) and the derived
  terrain streams. Two engine comments (fix.ts, drone.ts) that had
  inherited the upstream "mean bearings" / "expanding-square" wording are
  corrected — comments only, no code change; goldens unchanged.
• docs/upstream/DESIGN_NOTES.md re-synced to upstream 74b161a
  (SNAPSHOT.md updated; that pin is the design-notes copy, distinct
  from the fixtures' _meta.source_commit).

Changed

• CI: upstream-drift workflow bumped actions/checkout and
  actions/setup-node from v4 to v5 (Node 24 action runtime), clearing
  GitHub's Node 20 deprecation warning. The workflow still tests on
  Node 20, the package's stated minimum; setup-node v5's automatic
  package-manager caching does not engage because package.json has no
  packageManager field.
• CI: upstream-drift now also runs on every pull request and on push
  to main, so a PR whose engine no longer reproduces the committed
  fixtures, or whose fixtures no longer match live upstream fpv-sim
  main, shows a red check before it merges instead of failing the
  weekly run afterwards. The weekly schedule and manual dispatch remain
  for catching upstream movement while this repo is quiet. Paired with
  fpv-sim's new parity workflow (which regenerates these fixtures from
  a PR's index.html on that side), divergence between the two repos
  surfaces at merge time on whichever side changes.
• CI: the drift job is now a required status check on main
  (repository ruleset "main: require upstream-drift", no bypass actors),
  so a red upstream-drift run blocks the merge outright rather than
  merely flagging it. Repo setting, recorded here so it isn't invisible:
  renaming the job requires updating the ruleset's required context.

0.2.2 — Dependency hygiene (npm audit fix)

No engine or tool behavior changes; golden fixtures unchanged. Dependency
hygiene only.

Security

• npm audit fix (no --force): four transitive dependencies of
  @modelcontextprotocol/sdk bumped within their existing semver ranges
  to clear all reported advisories — hono 4.12.31 → 4.13.2 (ReDoS in
  CORS middleware, SSR memo() cross-request retention, proxy-helper
  Connection header handling, language-middleware complexity),
  @hono/node-server 1.19.14 → 1.19.17 (serve-static path traversal
  on Windows), fast-uri 3.1.4 → 3.1.5 (host confusion via backslash
  authority), ip-address 10.2.0 → 10.5.0 (leading-zero / CIDR /
  IPv4-mapped misclassification enabling SSRF bypasses). None of these
  code paths are exercised by this server (it uses node:http directly
  and the SDK's Streamable HTTP transport, no CORS/static/proxy
  middleware), so this is hygiene, not a fix for an exposure. SDK stays
  at 1.29.0; no dependency ranges changed. npm audit is clean
  afterwards.

0.2.1 — Version reporting fix + PolyForm relicense

No engine or tool behavior changes; golden fixtures unchanged.

Changed

• Relicensed from MIT to PolyForm Strict License 1.0.0
  (LICENSE.md): the repository stays public to read and
  use noncommercially, but is no longer open source — no modification,
  redistribution, or commercial use. Copies obtained under MIT before
  this change retain their MIT rights.

Fixed

• The version the server reports (MCP initialize handshake, /healthz,
  startup banner) is now read from package.json instead of a hardcoded
  constant in src/server/build.ts, which had silently stayed at 0.1.0
  through the 0.2.0 release — so the hosted demo's /healthz could not
  confirm a redeploy. package-lock.json had likewise never been bumped
  past 0.1.0; npm version now keeps it in step. A new test
  (test/version.test.ts) asserts the reported version equals
  package.json's.

Docs

• deploy/DEPLOY.md "Updating" now ends with a /healthz version check —
  the redeploy confirmation the fix above makes meaningful.
• Release tags v0.1.0 and v0.2.0 now exist (published 2026-08-16),
  so the compare links below use tag form. The [0.2.0] range
  previously ended one merge early (at PR #3 rather than the PR #4 merge
  that carried the version bump); corrected.

0.2.0 — Drift detection + Node 20 fix

Added

• Upstream drift detection (.github/workflows/upstream-drift.yml):
  weekly CI that regenerates the golden fixtures from live fpv-sim main
  and fails if engagement outcomes differ from the committed ones —
  distinguishing engine regressions (npm test) from upstream
  behavior changes (fixture comparison). Also warns when the pinned
  design-notes snapshot falls behind upstream.

Fixed

• npm test now works on Node 20, the stated minimum: the test glob is
  shell-expanded instead of relying on node --test pattern expansion,
  which only Node 21+ performs. Caught by the drift workflow's first CI
  run.

Docs

• README example transcript: added a postscript correcting the agent's
  closing guess that the 6-second launch stagger explained the residual
  BLUFOR edge — a 2,000-pair study on this engine
  (fpv-sim's MONTE_CARLO.md,
  experiment E2c) found the stagger moves outcome rates by less than half
  a point.

0.1.0 — Initial release

Added

• Initial release: headless TypeScript engine extracted from fpv-sim's
  index.html with float-exact behavior parity, proven by golden-master
  fixtures generated from the original browser source running in a Node
  vm (plus a real-browser cross-check).
• MCP server with five read-only, deterministic tools (run_engagement,
  sweep_seeds, compare_configs, describe_model, get_config_schema)
  and two documentation resources, with zod validation generated from the
  same parameter table as the schema tool.
• Stdio and Streamable HTTP transports, an MCP client demo
  (npm run demo), a VPS deployment runbook (deploy/), and the
  golden-fixture generator (scripts/generate-goldens.mjs).
• Headless-only additions to the sim's semantics: guaranteed termination
  (winner, both-drones-down, or 3600 s cap — the latter two reported as
  first-class STALEMATE outcomes) and flag timestamps for aggregation.