navicore/patch-prolog
1
0
Fork 0
Code Issues Pull requests Releases 11 Activity Actions

● M3 is complete — just ci fully green. Here's what landed: #1

Merged
navicore merged 1 commit from M3 into main 2026-06-04 19:59:45 +00:00
Conversation 0 Commits 1 Files changed 22 +3614 -388
navicore commented 2026-06-04 19:49:03 +00:00
Owner
Copy link

M3: Control, arithmetic, and indexing

Compiled control flow (codegen/lower.rs + body.rs): clause bodies now lower to a goal IR and compile to native
control —

  • ; → choice point retrying the right branch
  • (C -> T ; E) / once/1 → commit-height scratch slots in the body frame; the then-continuation truncates the CP
    stack
  • + G → a CP that continues the body (the driver's rewind undoes G's bindings for free); if G succeeds,
    cut-to-height and fail
  • ! → barrier captured at predicate entry, with correct ISO scoping: transparent through ;/->/,, local inside
    conditions, +, and once (a per-context cut-slot is threaded through compilation)

Builtins (background-agent port of v1's builtins.rs, 29 new runtime tests): full is/2 evaluable vocabulary,
comparisons, =/=/==/==/@-order/compare/3 — with error strings captured from the live v1 oracle and asserted
byte-for-byte (error(evaluation_error(zero_divisor), Division by zero (integer division)) etc.).

First-arg indexing: entry functions switch on the dereferenced first argument's tag/value to per-key clause
chains. Single-candidate dispatch pushes no choice point. The 5k-fact chain query went from ~10s (M2 linear
scan) to 3ms.

Query side: the runtime's minimal parser grew operator-precedence climbing + float literals, and a control
walker for runtime-built goals (walks goal terms only — never clauses; the v1 lesson holds). ./binary --query "X
is 2 + 3 * 4" works just like v1.

One finding worth your attention: the adversarial diff sweep exposed that v1 treats cut as opaque inside ; —
t(X) :- (m(X), X > 1, ! ; X = fallback) leaks fallback in v1, which violates ISO 7.8.4 (SWI/GNU give only X =
2). I chose ISO over bug-compatibility, documented it in ISO_COMPLIANCE.md under "Cut", and pinned both
behaviors' tests. Everything else in the sweep matches the oracle byte-for-byte, including linting.pl (which
needed +) end to end.

Verification: 71 runtime + 48 frontend + 14 M3-suite + 14 M2-suite + 11 golden-IR tests, cut-under-512KB-stack
recursion test, full just ci.

M3: Control, arithmetic, and indexing Compiled control flow (codegen/lower.rs + body.rs): clause bodies now lower to a goal IR and compile to native control — - ; → choice point retrying the right branch - (C -> T ; E) / once/1 → commit-height scratch slots in the body frame; the then-continuation truncates the CP stack - \+ G → a CP that continues the body (the driver's rewind undoes G's bindings for free); if G succeeds, cut-to-height and fail - ! → barrier captured at predicate entry, with correct ISO scoping: transparent through ;/->/,, local inside conditions, \+, and once (a per-context cut-slot is threaded through compilation) Builtins (background-agent port of v1's builtins.rs, 29 new runtime tests): full is/2 evaluable vocabulary, comparisons, =/\=/==/\==/@-order/compare/3 — with error strings captured from the live v1 oracle and asserted byte-for-byte (error(evaluation_error(zero_divisor), Division by zero (integer division)) etc.). First-arg indexing: entry functions switch on the dereferenced first argument's tag/value to per-key clause chains. Single-candidate dispatch pushes no choice point. The 5k-fact chain query went from ~10s (M2 linear scan) to 3ms. Query side: the runtime's minimal parser grew operator-precedence climbing + float literals, and a control walker for runtime-built goals (walks goal terms only — never clauses; the v1 lesson holds). ./binary --query "X is 2 + 3 * 4" works just like v1. One finding worth your attention: the adversarial diff sweep exposed that v1 treats cut as opaque inside ; — t(X) :- (m(X), X > 1, ! ; X = fallback) leaks fallback in v1, which violates ISO 7.8.4 (SWI/GNU give only X = 2). I chose ISO over bug-compatibility, documented it in ISO_COMPLIANCE.md under "Cut", and pinned both behaviors' tests. Everything else in the sweep matches the oracle byte-for-byte, including linting.pl (which needed \+) end to end. Verification: 71 runtime + 48 frontend + 14 M3-suite + 14 M2-suite + 11 golden-IR tests, cut-under-512KB-stack recursion test, full just ci.
● M3 is complete — just ci fully green. Here's what landed:
All checks were successful
CI - Linux / CI - Linux x86_64 (pull_request) Successful in 45s
Details
d9a64ffa26 
M3: Control, arithmetic, and indexing

  Compiled control flow (codegen/lower.rs + body.rs): clause bodies now lower to a goal IR and compile to native
  control —
  - ; → choice point retrying the right branch
  - (C -> T ; E) / once/1 → commit-height scratch slots in the body frame; the then-continuation truncates the CP
  stack
  - \+ G → a CP that continues the body (the driver's rewind undoes G's bindings for free); if G succeeds,
  cut-to-height and fail
  - ! → barrier captured at predicate entry, with correct ISO scoping: transparent through ;/->/,, local inside
  conditions, \+, and once (a per-context cut-slot is threaded through compilation)

  Builtins (background-agent port of v1's builtins.rs, 29 new runtime tests): full is/2 evaluable vocabulary,
  comparisons, =/\=/==/\==/@-order/compare/3 — with error strings captured from the live v1 oracle and asserted
  byte-for-byte (error(evaluation_error(zero_divisor), Division by zero (integer division)) etc.).

  First-arg indexing: entry functions switch on the dereferenced first argument's tag/value to per-key clause
  chains. Single-candidate dispatch pushes no choice point. The 5k-fact chain query went from ~10s (M2 linear
  scan) to 3ms.

  Query side: the runtime's minimal parser grew operator-precedence climbing + float literals, and a control
  walker for runtime-built goals (walks goal terms only — never clauses; the v1 lesson holds). ./binary --query "X
  is 2 + 3 * 4" works just like v1.

  One finding worth your attention: the adversarial diff sweep exposed that v1 treats cut as opaque inside ; —
  t(X) :- (m(X), X > 1, ! ; X = fallback) leaks fallback in v1, which violates ISO 7.8.4 (SWI/GNU give only X =
  2). I chose ISO over bug-compatibility, documented it in ISO_COMPLIANCE.md under "Cut", and pinned both
  behaviors' tests. Everything else in the sweep matches the oracle byte-for-byte, including linting.pl (which
  needed \+) end to end.

  Verification: 71 runtime + 48 frontend + 14 M3-suite + 14 M2-suite + 11 golden-IR tests, cut-under-512KB-stack
  recursion test, full just ci.
navicore referenced this pull request from a commit 2026-06-20 22:03:21 +00:00
PR #28 feedback addressed
navicore referenced this pull request from a commit 2026-07-03 13:59:38 +00:00
What landed
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
navicore/patch-prolog!1
No description provided.