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

● M4 is complete — full builtin parity with v1, just ci green. The headline numbers: #2

Merged
navicore merged 1 commit from M4 into main 2026-06-04 21:29:59 +00:00
Conversation 0 Commits 1 Files changed 41 +5679 -304
navicore commented 2026-06-04 21:28:29 +00:00
Owner
Copy link

M4 summary

The corpus verdict: v1's integration suite, ported to run against compiled binaries — 89 grouped tests covering
~200 of v1's 248 assertions — passes with zero ignores. Plus a permanent 57-goal differential corpus that runs
both implementations side by side (just diff-test, auto-skips in CI).

What landed:

  • Structured errors — relocatable ISO error balls that survive heap rewinding, byte-identical rendering to v1
  • catch/throw — catch frames as flagged choice points; the shared driver unwinds balls to the matching catcher;
    cut stops at catch frames; step limit stays uncatchable
  • findall/3 (bounded sub-search, nests, propagates errors), call/N + variable-goal metacall, between/3 (the one
    nondet builtin, dispatched like a compiled predicate)
  • All 25 v1 det builtins, oracle-verified — including quirks like compound([1]) being true and atom_concat being
    deterministic-only
  • Boxed i64 (TAG_BIG) — the i61 immediate limit is gone
  • v1's stdlib compiled into every binary — member, nondeterministic append, length, etc.

The triage story (your ISO-over-bug rule in action, but inverted): the corpus port surfaced 6 divergences, and
checking each against the live oracle showed all six were plgc bugs, not v1 bugs — so all six got fixed rather
than whitelisted:

  1. Query-walker cut had no barrier mechanics (+ (foo(X), !, ...) mis-pruned) → added qbarrier, snapshotted in
    every walker continuation frame, mirroring the compiled side's cut slots
  2. X = f(X) segfaulted the renderer → cycle-safe rendering with v1's exact cut-to-var behavior (f(_0))
  3. between(MAX, MAX, X) returned -1 → missing i64 boxing
  4. & 5. Prefix +/, standalone operator atoms (-), and the : operator missing from the runtime query parser →
    added
  5. One ported expectation was actually the porting agent's guess, contradicted by the live oracle — corrected to
    oracle truth

The oracle is now optional: semantics are pinned by the ported corpus itself, exactly as planned.

M4 summary The corpus verdict: v1's integration suite, ported to run against compiled binaries — 89 grouped tests covering ~200 of v1's 248 assertions — passes with zero ignores. Plus a permanent 57-goal differential corpus that runs both implementations side by side (just diff-test, auto-skips in CI). What landed: - Structured errors — relocatable ISO error balls that survive heap rewinding, byte-identical rendering to v1 - catch/throw — catch frames as flagged choice points; the shared driver unwinds balls to the matching catcher; cut stops at catch frames; step limit stays uncatchable - findall/3 (bounded sub-search, nests, propagates errors), call/N + variable-goal metacall, between/3 (the one nondet builtin, dispatched like a compiled predicate) - All 25 v1 det builtins, oracle-verified — including quirks like compound([1]) being true and atom_concat being deterministic-only - Boxed i64 (TAG_BIG) — the i61 immediate limit is gone - v1's stdlib compiled into every binary — member, nondeterministic append, length, etc. The triage story (your ISO-over-bug rule in action, but inverted): the corpus port surfaced 6 divergences, and checking each against the live oracle showed all six were plgc bugs, not v1 bugs — so all six got fixed rather than whitelisted: 1. Query-walker cut had no barrier mechanics (\+ (foo(X), !, ...) mis-pruned) → added qbarrier, snapshotted in every walker continuation frame, mirroring the compiled side's cut slots 2. X = f(X) segfaulted the renderer → cycle-safe rendering with v1's exact cut-to-var behavior (f(_0)) 3. between(MAX, MAX, X) returned -1 → missing i64 boxing 4. & 5. Prefix +/\, standalone operator atoms (-), and the : operator missing from the runtime query parser → added 5. One ported expectation was actually the porting agent's guess, contradicted by the live oracle — corrected to oracle truth The oracle is now optional: semantics are pinned by the ported corpus itself, exactly as planned.
● M4 is complete — full builtin parity with v1, just ci green. The headline numbers:
All checks were successful
CI - Linux / CI - Linux x86_64 (pull_request) Successful in 44s
Details
641bbf1204 
M4 summary

  The corpus verdict: v1's integration suite, ported to run against compiled binaries — 89 grouped tests covering
  ~200 of v1's 248 assertions — passes with zero ignores. Plus a permanent 57-goal differential corpus that runs
  both implementations side by side (just diff-test, auto-skips in CI).

  What landed:
  - Structured errors — relocatable ISO error balls that survive heap rewinding, byte-identical rendering to v1
  - catch/throw — catch frames as flagged choice points; the shared driver unwinds balls to the matching catcher;
  cut stops at catch frames; step limit stays uncatchable
  - findall/3 (bounded sub-search, nests, propagates errors), call/N + variable-goal metacall, between/3 (the one
  nondet builtin, dispatched like a compiled predicate)
  - All 25 v1 det builtins, oracle-verified — including quirks like compound([1]) being true and atom_concat being
  deterministic-only
  - Boxed i64 (TAG_BIG) — the i61 immediate limit is gone
  - v1's stdlib compiled into every binary — member, nondeterministic append, length, etc.

  The triage story (your ISO-over-bug rule in action, but inverted): the corpus port surfaced 6 divergences, and
  checking each against the live oracle showed all six were plgc bugs, not v1 bugs — so all six got fixed rather
  than whitelisted:
  1. Query-walker cut had no barrier mechanics (\+ (foo(X), !, ...) mis-pruned) → added qbarrier, snapshotted in
  every walker continuation frame, mirroring the compiled side's cut slots
  2. X = f(X) segfaulted the renderer → cycle-safe rendering with v1's exact cut-to-var behavior (f(_0))
  3. between(MAX, MAX, X) returned -1 → missing i64 boxing
  4. & 5. Prefix +/\, standalone operator atoms (-), and the : operator missing from the runtime query parser →
  added
  5. One ported expectation was actually the porting agent's guess, contradicted by the live oracle — corrected to
  oracle truth

  The oracle is now optional: semantics are pinned by the ported corpus itself, exactly as planned.
navicore referenced this pull request from a commit 2026-06-20 22:03:21 +00:00
PR #28 feedback addressed
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!2
No description provided.