poker-solver v1.8.0 — Cross-platform SIMD portability + terminal-utility convention purge + `.dmg` fork-bomb fix

Status: DRAFT (post-purge framed) — Phases 1-4 + AVX2 all merged to
main as of 77e751c (PR #32, 2026-05-26). The terminal-utility
convention purge lands as PR 78 (37e5be1)
and is the engine correctness fix in v1.8.0 — see "Engine fixes" §5
and "Migration / breaking changes" below. Engine + parity fixes from
the v1.7.1 bundle and v1.7.2 (.dmg fork-bomb fix + CI hardening) are
folded into v1.8.0 per docs/v1_6_1_ship_hold_review_2026-05-26.md
and docs/v1_7_1_tag_decision_2026-05-26.md. The v1.6.1 ship hold is
LIFTED under the canonical convention (see "Known issues remaining"
A83 entry). All substantive placeholders have been substituted; final
ship step is tag + GitHub release.

Release date: 2026-05-27
Tag: v1.8.0 (to be created at ship time)
Baseline commit on origin/main: eb74fb3 (PR #60, 2026-05-26),
plus the terminal-utility convention purge at 37e5be1
(PR 78).
Final tag SHA will be set at git tag time.

Headline

v1.8.0 — Cross-platform SIMD portability + terminal-utility convention purge + .dmg fork-bomb fix.

Three things land together in this release:

Cross-platform SIMD complete (portability win). The Discounted
CFR solver's hot inner loops are now hand-vectorized across NEON
(Apple Silicon), AVX2 + SSE2 (x86_64, runtime-detected), with an
automatic scalar fallback for anything else. This is a portability
win, not a speedup — measured wall-clock on Apple Silicon
(M4 Pro, aarch64) is within noise (~1.0×) because LLVM's -O3
autovectorizer already covers the small-slice case. The headline
value is closing the ~3-month gap between Apple Silicon and x86_64
coverage with a stable hand-written floor that doesn't depend on the
compiler's heuristics.
Terminal-utility convention purge (engine correctness fix). The
prior "rust" terminal-utility convention treated initial_contributions
as recoverable, which produced a per-action regret bias of 12-50pp
versus the reference Brown solver at deep cap. v1.8.0 adopts the
single canonical real-poker convention (winner collects the full pot
including dead money from prior streets). See "Engine fixes" and
"Migration / breaking changes" for the formula, a one-line numeric
example, and the rebaseline policy. Pre-v1.8.0 solver outputs are
NOT comparable to v1.8.0 outputs.
.dmg fork-bomb fix (CRITICAL). The v1.6.0 .dmg had an
uncontrolled-spawn bug on Finder launch
(multiprocessing.freeze_support() missing from the PyInstaller
entry point). v1.6.0 .dmg has been pulled from its GitHub Release;
v1.8.0 ships the repackaged build.

SIMD vector kernels (cross-platform): Explicit NEON/AVX2/SSE2
intrinsics replace the previous scalar inner loops in dcfr_vector.rs.
Bit-identical SIMD-vs-scalar output on the same convention; measured
wall-clock impact on Apple Silicon (M4 Pro, aarch64) is within noise
(~1.0×) because LLVM's autovectorizer at -O3 already covers the
small-slice case (action_count = 2-5 per decision row). Primary value
is portability (x86_64 with explicit AVX2 dispatch, runtime-detected)
and a stable hand-written floor that doesn't depend on the compiler's
heuristics. x86_64 wall-clock measurement is pending (no AVX2 hardware
in the bench fleet at time of write). Full benchmark report:
docs/v1_8_simd_perf_benchmark_2026-05-26.md.

Highlights

1. Cross-platform SIMD vector kernels (Phase 1-4)

Four hot loops in crates/cfr_core/src/dcfr_vector.rs are now
vectorized, each behind a single safe public API with runtime ISA
detection. Bit-identical output verified by an end-to-end cross-backend
smoke test that runs 1000 iterations of a Kuhn poker solve on both the
scalar and SIMD paths and asserts equality.

Phase	Kernel	PR	Merge SHA
1	`discount_regrets` + `discount_strategy_sum`	#23	`485aa8c`
1+	AVX2 runtime-detect path (x86_64)	#35	`db8d646`
2	`update_regret_sum`	#41	`8073bcc`
3	`update_strategy_sum`	#33	`a712950`
4	`compute_strategy`	#32	`77e751c`

Per-backend coverage:

NEON for target_arch = "aarch64" (Apple Silicon, ARM servers) —
128-bit, 2-lane f64.
AVX2 + SSE2 for target_arch = "x86_64", with
is_x86_feature_detected!("avx2") runtime dispatch:
- AVX2 path: 256-bit, 4-lane f64 (Haswell+, ~2013 and later).
- SSE2 baseline: 128-bit, 2-lane f64 (any x86_64 CPU).
Scalar fallback for any other architecture and for the epilogue
(the 0-3 trailing elements that don't fit a SIMD lane).

Why this matters: the previous v1.7.x line was already NEON-optimized
for Apple Silicon, but x86_64 users (Intel Macs, Windows PCs, Linux
desktops, CI runners) got the scalar fallback. That's a ~3-month gap
between Apple Silicon and x86_64 performance, now closed (at the
architecture level — the empirical x86_64 wall-clock measurement
remains pending). The cross-platform spec
(docs/pr_proposals/v1_8_cross_platform_simd_spec.md) drives the
architecture. Sarah persona W2.3 remains pending the post-v1.8 retest
(turn-fixture, agent in flight 2026-05-26); the original "unblocked
on M-series" projection was tied to a 4-8× SIMD speedup that did not
materialize on M4 Pro arm64 (see "Persona test status" below).

2. `.dmg` fork-bomb fix (CRITICAL)

The v1.6.0 .dmg had a critical fork-bomb on Finder launch.
Double-clicking Poker Solver.app on macOS spawned the app's
multiprocessing workers recursively because
multiprocessing.freeze_support() was missing from the PyInstaller
entry point. Each spawned child re-launched the parent process,
exponentially — freezing the user's Mac.

Resolution:

PR #42 (728206e) adds the multiprocessing.freeze_support()
guard to the PyInstaller entry point, eliminating the fork-bomb on
the v1.8.0 build.
The v1.6.0 .dmg asset has been retroactively pulled from its
GitHub Release. The v1.6.0 release page now carries a critical
warning pointing users at the v1.8.0 repackaged build (or the
from-source install) instead.
A user-facing warning was also added to the CHANGELOG retroactive
amendment for v1.6.0.

If you previously downloaded the v1.6.0 .dmg, delete the
Poker Solver.app bundle and switch to the v1.8.0 .dmg when this
release publishes. If you ran the v1.6.0 .dmg and your Mac froze on
launch, no persistent damage was done — the workers stopped when the
parent process was killed.

Full RCA: docs/dmg_spawn_loop_rca_2026-05-26.md.

3. Lint / format / deps green-up (PR #43)

PR #43 (cfc6bc5) is a non-functional cleanup pass that brings
the main-branch lint gates back to green after the v1.7.x churn:

cargo clippy --all-targets --release — clean.
ruff check . — clean (lint).
ruff format --check . — clean (format). black is removed in favor
of ruff format; the two formatters had diverged on a handful of
multi-line argument lists.
rich>=13.0 added as an explicit runtime dependency. It was
implicitly required by poker-solver parity (Brown reference parity
CLI subcommand) via rich.console.Console imports, but the install
wouldn't fail until users actually invoked the subcommand.

No public API or algorithm changes.

4. Documentation accuracy (PRs #44, #45)

Two doc-only PRs land alongside the SIMD bundle:

PR #44 (a6b89f7) fixed three executable code bugs in
README.md and USAGE.md examples:
1. Range parser example used the wrong constructor signature.
2. A solve_river_nash() example was missing the iterations=
  keyword.
3. A hero/board overlap example assumed a card slot that the API
  rejects.
  Plus a stale claim about "no CLI for parity comparison" was scrubbed
  (the poker-solver parity subcommand has shipped since v1.7.0).
PR #45 (dbfc8d0) cleaned up downstream doc drift:
- USAGE.md header version stamp brought to v1.7.x.
- CHANGELOG.md .dmg retroactive amendment for v1.6.0 (see
  Highlight 2).
- docs/dmg_install_guide.md banner pointing to the v1.8.0 repackaged
  build.
- Supersede banners on a handful of stale planning docs.

5. Engine + parity-wrapper fixes carried from the v1.7.1 bundle

v1.8.0 inherits a 10-PR bundle of engine and parity-wrapper fixes that
landed piecewise on main between v1.7.0 and v1.8.0. No formal
v1.7.1 tag was created (per docs/v1_7_1_tag_decision_2026-05-26.md);
the fixes are folded into this release. The v1.6.1 ship hold has been
lifted per docs/v1_6_1_ship_hold_review_2026-05-26.md.

Engine correctness:

PR 78 (37e5be1) — Terminal-utility
convention purge (canonical real-poker convention). Prior versions
used a "rust" terminal-utility formula that treated
initial_contributions as recoverable by the player who folded; this
produced a per-action regret bias that diverged 12-50pp from the
reference Brown solver at deep cap (PR #93 ablation,
docs/a83_terminal_utility_ablation_results_2026-05-26.md). v1.8.0
deletes that path and ships a single canonical utility() function.

Canonical formula (per feedback_brown_convention_adopt.md):
```
winner_utility = pot_total - contrib_subgame_winner
               = base_pot + contrib_subgame_loser     (in BB)
loser_utility  = -contrib_subgame_loser               (in BB)
tie_each       = pot_total/2 - contrib_subgame_player (in BB)
```
where contrib_subgame[i] = state.contributions[i] - cfg.initial_contributions[i]
and pot_total = cfg.initial_pot + contrib_subgame[0] + contrib_subgame[1].

One-line numeric example. At a canonical leaf where P0 folds with
base_pot = 10 BB and zero in-subgame contributions
(contrib_subgame = (0, 0), initial_contributions = (5, 5)):
the old "rust" path returned (-5, +5) BB (treating each player's
seed-split half-pot as still recoverable); the canonical path
returns (0, +10) BB — P1 (the non-folder) collects the full pot
including the dead money already on the table, P0 loses only their
in-subgame contribution (zero here).

The game is constant-sum under the canonical convention
(sum of utilities = +base_pot/bb per leaf rather than zero); DCFR
convergence is unaffected (proofs require bounded utilities + finite
action set, not zero-sum). No feature flag, no
TerminalUtilityConvention enum, no runtime switch — one formula.
See "Migration / breaking changes" below for the rebaseline policy.
PR 50 (#5) — Phantom ALL_IN action menu guard (paired Rust +
Python). At facing-all-in nodes, the responder's action menu no
longer emits a degenerate ALL_IN raise that would have had no chip
distinction from calling. Closes R6.
PR 51 / PR #16 — dcfr_vector.rs off-by-one + asymmetric-range
next_reach sizing fix; closes the asymmetric-range solve panic.

Brown parity test harness — wrapper boundary corrections:

PR 52 (#8) — Suit-encoding char-to-char fix; replaced a silent
paired h ↔ d swap with explicit mapping. Closes R8.
PR 55 (#10) — P0/P1 player-index convention swap at the wrapper
boundary. Closes R9.
PR 56 (#12) — Hand-string sort-order canonicalization at the
wrapper boundary. Closes R10.

Brown parity test harness — renderer + acceptance reframe:

PR 54 (#9) — Renderer stack_ceiling kwarg + "A" token for
bets/raises at stack ceiling.
PR 53b (#14) + PR 53c (#15) — Brown apples-to-apples acceptance
test reframed from strict 5e-2 per-action gate to a 4-layer gate
(L1 structural / L2 shallow-strict / L3 deep-directional L1 ≤ 1.9 /
L3' p75 L1 ≤ 0.60 / L4 top-action ≥ 60%). Codifies that external
reference solvers (Brown, Pluribus) are sanity checks rather than
strict ground truth when action menus differ and deep-cap subgames
have indifference manifolds. PASSES on both river spots under
Dry-run #10 (2026-05-24) — see docs/v1_6_1_dryrun_10.md.

Ship-hardening + CI:

PR 59 (#18) — memory_profiler golden-file refresh + regen-mode
flag.
PR 60-equivalent (folded into PR #22) — _skip_or_fail() helper
- STRICT_ACCEPTANCE=1 env var: Brown parity test hard-fails on
  missing prereqs in strict mode.

The bundle resolves the 22-42pp Brown apples-to-apples deep-cap
divergence reported in v1.5.0 / v1.5.1 / v1.6.0 dry-runs as a
combination of test-side wrapper bugs (R8/R9/R10), one engine-side
mechanical guard (R6, PR 50), the terminal-utility convention
purge (PR 78, the engine correctness fix; see the
first bullet under "Engine correctness" above), and Nash multiplicity
on the residual at deep-cap indifference manifolds (see "Known issues
remaining" below). Pre-purge framings that treated the convention as
a "documented Brown design divergence" are explicitly retracted per
feedback_brown_convention_adopt.md; there is one canonical
convention.

6. v1.7.2 entries folded into v1.8.0

Per the v1.7.1 tag decision and the v1.6.1 ship-hold review, the
following v1.7.2 entries are also folded into v1.8.0 rather than
shipping under a separate tag:

PR #21 (bed37c4) — CI-driven release workflow (PR 62 origin).
PR #22 (1fefaff) — ship-process hardening Guards B + C
(PR 65 origin). Includes _skip_or_fail() + STRICT_ACCEPTANCE=1.
PR #42 (728206e) — .dmg fork-bomb fix
(multiprocessing.freeze_support()); the v1.6.0 .dmg was pulled
retroactively. See Highlight 2 for the user-facing impact.

CHANGELOG follows Option A alignment (fold v1.7.2 entry into
v1.8.0, do not preserve a separate v1.7.2 section) per
docs/v1_8_0_release_notes_prep_2026-05-26.md Open Question #1.

7. Documentation + ship-process cleanup wave

A user-impactful subset of the documentation / packaging / CI cleanup
PRs that landed on main between v1.7.0 and v1.8.0 (in addition to
the lint/clippy/deps green-up and doc-accuracy PRs already called out
in §3 and §4 above):

PR #46 — mypy resolves 7 substantive type errors
(post-cleanup follow-up; tightens runtime correctness for the public
Python API).
PR #47 — .dmg packaging hardening: arch label + version
stamp accuracy follow-up to PR #42's fork-bomb fix. Removes a
user-confusing "universal" arch label on a .dmg that ships an
arm64-only _rust.so.
PR #48 — USAGE.md §7b header refresh (v1.4.x → v1.7.x)
with regime guidance (salvaged from a stale PR #24).
PR #54 — v1.8.0 release execution script + .dmg
build runbook (codifies the ship sequence so the next release does
not relitigate the build-environment quirks).
PR #58 — CHANGELOG note for the poker-solver PATH shim
quirk (see "Known issues remaining" below for the user-facing
workaround).
PR #59 — persona-test status snapshot refresh
post-W3.2 / W3.4 retests (see "Persona test status" below).
PR #60 — track previously-orphan reference targets
(closes 7 dangling-reference lints from the doc-cross-check sweep).

This cleanup wave does not change any runtime behavior; it lifts the
release boundary to a clean lint / docs / packaging baseline.

8. Persona test status (post-v1.8 SIMD + W3.2 + W3.3 + W3.4 retests)

The persona-test framework's counts as of 2026-05-26 (post the W3.2 +
W3.3 + W3.4 retests; W2.3 turn-fixture retest IN PROGRESS):

Verdict	Count	Workflows
PASS	10	W1.1, W1.2, W1.3, W1.4, W2.5, W3.1, W3.2 (new), W3.3 (new; P2 closing test), W3.4 (caveated), W4.1, W4.3 (W4.3 via aggregator path)
PARTIAL	4	W1.5, W2.1, W2.2, W2.4, W4.2
BLOCKED	2	W2.3 (pending retest), others closed by v1.8 + PR #38
FAIL	1	W3.5 (Type B-DOC; docs-only follow-up, no code patch needed per `v1_7_1_wrapper_fix_spec.md`)

Net delta vs the 2026-05-25 snapshot (7 / 5 / 4 / 2): +3 PASS, -2
BLOCKED, -1 FAIL→PARTIAL, +1 PARTIAL→PASS (W3.3 PARTIAL → PASS via
PR #66 P2 closing test for node-locking-at-scale; W3.5 reclassified to
PARTIAL-Type-B-DOC after the v1.8 retest confirmed bit-identical-to-v1.7.0
behavior).

W3.2: BLOCKED → PASS. PR 76 (PR #38, commit feee974) shipped
solve_best_response(game, opponent_strategy, *, hero_player) plus
a poker-solver best-response CLI subcommand; the Kuhn smoke shows
exploit_gap_bb > 0 on both seats. Type A (correctness; cleared the
named API blocker). Source: docs/persona_w3_2_smoke_2026-05-26.md.
W3.3: PARTIAL → PASS. PR #66 (P2 closing test for
node-locking-at-scale) — all 4 acceptance criteria PASS (lock
passthrough bit-exact <1e-9; villain L1 = 0.3070 at facing-raise
node; EV invariant 5.0/5.0 BB on indifference manifold; 5 downstream
infosets diverge >1%); 3.00 s Python wall-clock (well under Daniel's
15 min budget). Type A (correctness; node-locking primitive verified
at scale). Source: docs/persona_w3_3_retest_2026-05-26.md.
W3.4: BLOCKED → PASS (caveated). PASS on the repurposed
monotone-river 3-bet-pot polarization fixture (80.71 s wall-clock,
27% of the Sarah 300 s gate); all 7 acceptance thresholds met.
The W3.4 unblock was via fixture-repurposing, NOT via the v1.8
SIMD perf gain — the original W3.4 flop MDF fixture remains
perf-bound; v1.8 SIMD measured ~1.0× on M4 Pro arm64 did not deliver
the projected 4-8× speedup. Source:
docs/persona_w3_4_retest_2026-05-26.md.
W3.5: FAIL → PARTIAL (Type B-DOC). v1.8 SIMD bit-identical to
v1.7.0 on the 6-class smoke (delta ~0); classification stands per
v1_7_1_wrapper_fix_spec.md (docstring + regression-test gaps; no
code patch needed). Source:
docs/persona_w3_5_retest_2026-05-26.md.
W2.3: PENDING. Post-v1.8 retest agent in flight (a99ec2e) on
the pre-staged turn fixture. Final tally will be 10 or 11 PASS / 1
or 2 BLOCKED depending on the wall-clock measurement.

Full per-workflow status: docs/persona_test_status_2026-05-26.md
(supersedes the 2026-05-25 baseline). The post-v1.8 audit framing is
captured at docs/persona_status_post_v1_8_2026-05-26.md.

Note on persona baselines vs. the convention purge. Persona tests
whose acceptance gate depends on absolute strategy probabilities,
exploitability magnitudes, or numerical agreement with a pre-v1.8.0
baseline need re-collection under the canonical convention (see
"Migration / breaking changes" below). Persona tests gated on
structural / topological / wall-clock criteria (W3.4 wall-clock,
W3.2 API existence, etc.) are unaffected by the convention purge.
The bit-identical-to-v1.7.0 note for W3.5 above refers to the
pre-purge SIMD smoke (SIMD-vs-scalar, same convention); under the
canonical convention, all v1.7.x baselines are stale.

Known issues remaining

The following are NOT resolved in v1.8.0 and remain open for future
work:

A83: Deep-cap Brown apples-to-apples residual gap (post-convention-purge)

A83 deep-cap residual (Nash multiplicity, not bug): Post-purge v1.5
Brown apples-to-apples test PASSES on both K72 + A83 spots under the
reframed 4-layer SANITY gate. Strict-per-cell |Δ| values: K72 0.852,
A83 0.907 — these reflect Nash multiplicity at deep-cap indifference
manifolds (Brown picks one Nash, our solver picks another; both are
valid). The reframed gate does NOT strict-assert per-cell (that would
be non-falsifiable under multiplicity per
feedback_nash_multiplicity_acceptance.md); it gates on L1, top-action
coverage, and overall coverage. The convention purge (PR #78,
37e5be1) closed the regret-update bias (PR #93's measured 12-50pp
Rust-vs-Rust shift); residual is genuine multiplicity. EV-of-action
invariance gauntlet (docs/ev_invariance_sanity_gauntlet_design_2026-05-27.md)
is the canonical sanity check for the residual.

Status post-purge: L1 max 1.703 (K72) / 1.813 (A83); strict max |Δ|
0.852 / 0.907; both fixtures PASS the reframed 4-layer SANITY gate
(strict per-cell layer is informational, not asserted, per
feedback_nash_multiplicity_acceptance.md) (v1.5 Brown
apples-to-apples residual after the canonical-convention adoption,
captured 2026-05-27 from the post-purge dry-run; full numbers at
docs/v1_5_brown_post_purge_numbers_2026-05-27.md).

Whatever remains after the convention purge is Nash multiplicity at
indifference manifolds rather than a game-definition mismatch: deep-cap
HUNL has indifference manifolds where multiple Nash equilibria are all
valid and small initial-condition perturbations select between them.
The convention purge DID fix the regret-update bias (PR #93's 12-50pp
gap is closed); the strict per-cell |Δ| residual is a SEPARATE
phenomenon and is NOT expected to collapse — A83's strict max |Δ|
moved from 0.907 (pre-purge) to 0.907 (post-purge) unchanged at the
recorded precision. The test PASSES because the reframed 4-layer gate
does not strict-assert per-cell. This is testable via EV-of-action
invariance — for a constant-sum game, the EV of each action at each
infoset is unique across all Nash equilibria, even when strategy
probabilities aren't (see
docs/ev_invariance_sanity_gauntlet_design_2026-05-27.md).

Trajectory of 0.907 (max |Δ| on the A83
fixture at iter=2000) anchors the post-purge baseline relative to
the pre-purge 12.27pp @ 2000 iters / 10.28pp @ 8000 iters figure
from PR #93's ablation. Note: the 12-50pp PR #93 gap was a
Rust-vs-Rust ablation between conventions and IS closed by the purge;
the 0.907 strict max |Δ| is the Rust-vs-Brown apples-to-apples residual
under the canonical convention, dominated by Nash multiplicity.

v1.6.1 hold-lift validated. The hold-lift decision stands under
the canonical convention: the v1.5 Brown apples-to-apples reframed
4-layer gate (L1 structural / L2 shallow-strict / L3 deep-directional /
L4 top-action ≥ 60%) PASSES on both river spots under Dry-run #10
(2026-05-24), and the post-purge residual is Nash-multiplicity rather
than a bug. Pre-purge premature framings (PR #49 "fully closed via
Nash multiplicity," the arbitrator NOT-A-BUG verdict) are explicitly
retracted per feedback_brown_convention_adopt.md and the supersede
banners landed by PR #75.

Gate 4 200K-iter river/turn validation results — PROVISIONAL,
re-run pending. The 2026-05-25 Gate 4 200K river-phase run
(reported 5.28e-14 mbb/g exploitability, monotone clean) and the
2026-05-26 Gate 4 200K v2 attempts (river: 186-byte empty-strategy
output; turn: killed mid-run at ~60 min) both used the same
solve --hunl-mode postflop --backend rust CLI path without
--initial-hole-cards, which routes through
_rust.solve_hunl_postflop with HUNLConfig.initial_hole_cards = None. Per docs/a83_track_a_results_analysis_2026-05-26.md §2c,
HUNLState::chance_outcomes returns Vec::new() defensively when
hole_cards = None, the scalar CFR loop's chance branch iterates
over the empty list and returns [0.0, 0.0] immediately, no
infoset is ever inserted into strategy_sum, and the reported
exploitability is then recomputed by _rust.compute_exploitability
against an empty strategy map (falls back to uniform(n_actions)
on every cache miss per exploit.rs:472-475). The earlier
5.28e-14 figure was from the prior nohup that may have hit the
same no-op path; treat that value as PROVISIONAL until a re-run
via the correct entrypoint (or via solve_range_vs_range_nash)
is completed. The Gate 4 200K validation does NOT block the
v1.8.0 release boundary: the v1.5 Brown apples-to-apples
acceptance test (Dry-run #10) is the canonical pass; Gate 4 was
a secondary high-iter sanity check.
Silent no-op hardening (resolved, PR #69 / 98fb503). PR #69
hard-fails the solve --hunl-mode postflop --backend rust CLI when
--initial-hole-cards is missing; previously this path silently
returned empty strategies (root cause: chance_outcomes() returned
Vec::new() defensively when hole_cards = None, so the scalar
CFR loop iterated over zero outcomes and strategy_sum stayed empty
across all iterations — the exploitability recompute then walked a
near-empty tree with uniform-fallback strategy cache and reported
bogus numbers). Two surgical additive validation checks (~5 LOC each)
close the asymmetry with the vector-form path (which already rejected
the inverse case at dcfr_vector.rs:806-811); no behavior change for
callers already supplying initial_hole_cards. This caused at least
one false-positive multi-hour investigation (Track A nohup, ~2 hours
of agent time on a zero-output experiment) before the RCA landed.
See docs/chance_outcomes_empty_rca_2026-05-26.md.
Apple Silicon arch hazard (dev-environment, pre-existing).
pyenv's x86_64 Python can't load the arm64 _rust.so; use
.venv/bin/python to avoid silent SKIPs. Symptoms: silent SKIPs
in pytest output for parity / diff tests, ModuleNotFoundError
on poker_solver._rust, or dlopen "missing symbol" errors when
the underlying issue is actually an arch mismatch. See CONTRIBUTING.md
- docs/poker_solver_shim_fix_2026-05-26.md.
poker-solver PATH shim quirk (dev-environment, pre-existing).
If poker-solver on PATH fails with ModuleNotFoundError: No module named 'poker_solver', the shim is likely resolving against a stale
editable-install .pth from a deleted temp worktree in another
Python on PATH. Two equally good workarounds:
- ./.venv/bin/poker-solver ... (from project root)
- python -m poker_solver.cli ... (with .venv activated)
  Cleanup: pip uninstall poker_solver from the broken Python env,
  then pip install -e . from project root. This is a pre-existing
  dev-environment quirk surfaced by W3.2 smoke — NOT a v1.8 regression.
  Full diagnostic: docs/poker_solver_shim_fix_2026-05-26.md. The
  CHANGELOG carries the same workaround text under v1.8.0 known issues
  (PR #58).
Range fractional frequencies. The spec for true fractional
frequencies (e.g. AKs:0.6) landed in PR #36; the
implementation is tracked separately and not in v1.8.0.
.app / .dmg notarization. The v1.8.0 .dmg is still
ad-hoc-signed (same as v1.6.0). Apple Developer enrollment is a
user carry-item and gates full notarization.

Migration / breaking changes

Terminal-utility convention change (BREAKING for output comparability)

Pre-v1.8.0 solver outputs are NOT comparable to v1.8.0 outputs.
v1.8.0 adopts the canonical real-poker terminal-utility convention
(see "Engine fixes" §5 above for the formula and a one-line numeric
example). The prior "rust" convention treated initial_contributions
as recoverable by the player who folded; the canonical convention does
not. The two conventions produce different strategy probabilities
even on identical seeds + identical action menus — the per-action
regret bias measured between conventions was 12-50pp at deep cap
(PR #93 ablation,
docs/a83_terminal_utility_ablation_results_2026-05-26.md).

Justification: per feedback_brown_convention_adopt.md, there is
exactly one correct terminal-utility convention — the rule of real
poker, in which the winner collects every chip in the pot including
dead money already on the table from prior streets. The "rust"
convention was an implementation error; it is not a feature flag, not
a runtime switch, and not a valid alternative. No
TerminalUtilityConvention::Rust variant exists in v1.8.0.

Rebaseline policy. Any artifact captured under the pre-v1.8.0
convention must be regenerated against v1.8.0:

Exploit-diff snapshots. Re-run against v1.8.0; do not compare to
v1.7.0-and-earlier exploit numbers numerically. The class of game
changed from zero-sum to constant-sum (sum of utilities =
+base_pot/bb per leaf rather than zero), which moves the
exploit-vs-Nash absolute number even when the strategies are
qualitatively the same.
Persona test baselines. Persona retest wrappers re-baseline
against v1.8.0; pre-v1.8.0 PASS / FAIL counts for any persona test
whose acceptance gate depends on strategy probabilities or
exploitability magnitudes are stale and need re-collection.
Fixtures + golden files. Any deterministic fixture or
golden-file artifact captured under the pre-v1.8.0 convention is
regenerated under v1.8.0. Don't preserve them via a flag; they were
captured under an incorrect game model.

If you have downstream tooling that diff-tests v1.8.0 against a
pre-v1.8.0 baseline expecting numerical agreement, that test will fail
by design — replace the baseline with a v1.8.0 capture.

Upgrade path

From v1.7.0 source install

git pull
pip install -e .
maturin develop --release

The Rust binary _rust.cpython-313-darwin.so (or the equivalent on
your platform) must be rebuilt because the SIMD kernels are new
Rust source. pip install -e . alone is not enough; you need
maturin develop --release (or pip install -e . --force-reinstall
which triggers maturin rebuild).

From the v1.6.0 `.dmg`

Delete the existing Poker Solver.app bundle (Finder → drag to
Trash).
Download the v1.8.0 .dmg from the GitHub Release page (when
published — note: the v1.8.0 .dmg build verification is the
final ship step before this release publishes).
Drag-install as usual.

The v1.6.0 .dmg was pulled; do not attempt to redownload it.

Compatibility

No public API changes. All vectorization is behind the existing
Rust extension's safe Python API.
CFR algorithm: bit-identical SIMD-vs-scalar on the same
convention. The cross-backend SIMD smoke test asserts bit-identical
regret + strategy_sum tables across NEON / AVX2 / SSE2 / scalar at
every iteration.
CFR output vs v1.7.0: NOT bit-identical because of the
terminal-utility convention purge (see "Migration / breaking
changes" above). The change is in the game definition (terminal
payoff at winner-take-all leaves), not in the CFR update rule.
Min Python: unchanged from v1.7.0.
Min Rust toolchain: stable, unchanged from v1.7.0.
crates/cfr_core version bump: 0.7.x → 0.8.0.
black is no longer a dev dependency. ruff format is the
formatter. If your local dev environment installed black
explicitly, you can leave it or pip uninstall black — ruff format is run by the project's pre-commit and CI gates.

If you observe non-bit-identical output between two v1.8.0 backends
(NEON vs AVX2 vs SSE2 vs scalar) on the same problem, please open an
issue — that is a regression by definition. Non-identical output
between v1.7.0 and v1.8.0 is expected per the convention purge.

Verification

Run the cross-backend smoke test yourself:

cargo test --release --features simd -- --nocapture cross_backend_smoke

This solves Kuhn poker for 1000 DCFR iterations on both the scalar and
SIMD paths and asserts bit-identical regret + strategy_sum tables at
every iteration.

To verify the AVX2 dispatch on x86_64 with an AVX2-capable CPU:

RUST_LOG=info cargo test --release --features simd -- compute_strategy_avx2

Full PR list

All PRs that ship in v1.8.0, in merge order:

#23 485aa8c — feat(simd): v1.8 Phase 1 — cross-platform discount kernel
#35 db8d646 — feat(simd): v1.8 AVX2 runtime-detect for x86_64
#41 8073bcc — feat(simd): v1.8 Phase 2 — update_regret_sum
#33 a712950 — feat(simd): v1.8 Phase 3 — update_strategy_sum
#32 77e751c — feat(simd): v1.8 Phase 4 — compute_strategy
#42 728206e — fix(dmg): freeze_support() patches v1.6.0 fork-bomb + warn users
#43 cfc6bc5 — chore: green up lint/clippy/format/deps gates on main
#44 a6b89f7 — docs: fix executable code bugs in README + USAGE
#45 dbfc8d0 — docs: drift cleanup v2 (USAGE header, CHANGELOG .dmg, supersede banners)
#78 37e5be1 — fix(engine): terminal-utility convention purge (canonical real-poker convention; single utility() function; deletes "rust" convention path)

Plus the CI-hardening prerequisites that landed earlier in the v1.7.x
sequence but enable the cross-platform SIMD CI matrix:

#20 — feat(ci): cross-platform CI matrix (Ubuntu x86_64 + macOS arm64)
#21 bed37c4 — PR 62: v1.7.2 — CI-driven release workflow
#22 1fefaff — PR 65: ship-process hardening Guards B + C
#28 6e9fb11 — fix(scripts): auto-bootstrap references when build_noambrown runs on fresh checkout
#30 7edb1aa — test(cfr_core): v1.8 cross-platform end-to-end SIMD smoke

Acknowledgments

This release closes a ~3-month gap between Apple Silicon and x86_64
performance and a ~2-month-old critical packaging bug. Thanks to:

The cross-platform SIMD spec
(docs/pr_proposals/v1_8_cross_platform_simd_spec.md) for the
original architecture plan.
The cherry-pick / merge-orchestration workflow for letting the 4
SIMD phases land cleanly across the v1.7.x churn without merge
conflicts.
The persona-test framework (Sarah W2.3, Marcus turn-spot) for
empirically anchoring the performance budgets that motivated the
release.

What's next

v1.8.1+: vectorize secondary hot paths (range expansion,
payoff computation at terminal nodes) if profiling shows they're now
the dominant cost.
v1.9 candidate: EMD bucketing for flop interactive viability.
Since v1.8 SIMD measured ~1.0× on M4 Pro arm64 (root cause: LLVM
-O3 already autovectorizes the small-slice case), EMD bucketing
is now the more likely lever for unblocking the perf-bound turn /
flop persona workflows (W2.3 / W3.4 flop / W2.1 / W2.4); the W2.3
retest in flight will set the final wall-clock baseline.
v2.0: full-tree preflop solver + range-based dealing (currently
on a long-running v2 feature branch).

Navigation Menu

Repository navigation

v1.8.0 — Cross-platform SIMD + .dmg fork-bomb fix

Choose a tag to compare

Sorry, something went wrong.

Sorry, something went wrong.

Uh oh!

No results found

poker-solver v1.8.0 — Cross-platform SIMD portability + terminal-utility convention purge + `.dmg` fork-bomb fix

Headline

Highlights

1. Cross-platform SIMD vector kernels (Phase 1-4)

2. `.dmg` fork-bomb fix (CRITICAL)

3. Lint / format / deps green-up (PR #43)

4. Documentation accuracy (PRs #44, #45)

5. Engine + parity-wrapper fixes carried from the v1.7.1 bundle

6. v1.7.2 entries folded into v1.8.0

7. Documentation + ship-process cleanup wave

8. Persona test status (post-v1.8 SIMD + W3.2 + W3.3 + W3.4 retests)

Known issues remaining

A83: Deep-cap Brown apples-to-apples residual gap (post-convention-purge)

Migration / breaking changes

Terminal-utility convention change (BREAKING for output comparability)

Upgrade path

From v1.7.0 source install

From the v1.6.0 `.dmg`

Compatibility

Verification

Full PR list

Acknowledgments

What's next

Uh oh!

v1.8.0 — Cross-platform SIMD + .dmg fork-bomb fix

poker-solver v1.8.0 — Cross-platform SIMD portability + terminal-utility convention purge + .dmg fork-bomb fix

Headline

Highlights

1. Cross-platform SIMD vector kernels (Phase 1-4)

2. .dmg fork-bomb fix (CRITICAL)

3. Lint / format / deps green-up (PR #43)

4. Documentation accuracy (PRs #44, #45)

5. Engine + parity-wrapper fixes carried from the v1.7.1 bundle

6. v1.7.2 entries folded into v1.8.0

7. Documentation + ship-process cleanup wave

8. Persona test status (post-v1.8 SIMD + W3.2 + W3.3 + W3.4 retests)

Known issues remaining

A83: Deep-cap Brown apples-to-apples residual gap (post-convention-purge)

Migration / breaking changes

Terminal-utility convention change (BREAKING for output comparability)

Upgrade path

From v1.7.0 source install

From the v1.6.0 .dmg

Compatibility

Verification

Full PR list

Acknowledgments

What's next

Uh oh!

poker-solver v1.8.0 — Cross-platform SIMD portability + terminal-utility convention purge + `.dmg` fork-bomb fix

2. `.dmg` fork-bomb fix (CRITICAL)

From the v1.6.0 `.dmg`