Skip to content

Forward merge release/v0.4 (a1b925dc6: v0.4.0, H5, H6, the release-workflow fixes) into release/v0.5 - #327

Merged
afflom merged 82 commits into
release/v0.5from
claude/forward-merge-v0.4-v0.5-2
Oct 27, 2026
Merged

afflom merged 82 commits into
release/v0.5from
claude/forward-merge-v0.4-v0.5-2

Conversation

@afflom

@afflom afflom commented Oct 27, 2026

Copy link
Copy Markdown
Member

What

The second forward merge of release/v0.4 into release/v0.5 (SPEC.md §9.4), in the form of #322: a branch cut from release/v0.5 (61b4638cd), release/v0.4 at a1b925dc6 merged into it by merge commit, the conflicts resolved, the release/v0.4 specification text ported into the chapter modules of spec/, and the identities regenerated on the merged tree.

It brings the v0.4.0 tag commit e0b74b81e and the first post-tag landing onto this line: the heap half's H5 (#298) and H6 (#304) of #91 stage 4 and the release-workflow fixes (#320, #321, #323).

The merge

  • Generated files (locks, expected/, packages, provenance, fixtures, records) took the release/v0.4 side and were regenerated here.
  • Hand-owned conflicts, three-way: both RB-21 (the runtime text, from release/v0.4) and RB-22 (the loop ends of Production: a looped fallible function cannot tail-call another fallible function (LLB6005) #279, renumbered at Forward merge release/v0.4 (0.4.0, 93fbe7df9) into release/v0.5 #322) are kept in the dispatch, the named tests, and the feature; the bootstrap validator counts are the sum of both sides' deltas over the common base (37 loop crates and 44 machine families; 473 runs of validateCrate, 216 refused; 336 of correspondCrate, 103 refused); the register count stays projected (NS-03), not a literal; spec_links.rs and repository.rs keep this line's anti-vacuity form.
  • Specification: on this line SPEC.md and model/ids.toml are projections of spec/src, so the release/v0.4 text was ported with cargo xtask spec port (25 blocks of Chapter09, Chapter17 and Chapter30) and two insert-after calls (the SP-30 and RB-21 paragraphs); the rows RB-21 and SP-30 were added to IdsRB/IdsSP, and RB-16, RB-17, RB-19, RB-20, SP-29, BS-05, BS-06, BS-09 take their 0.4.0 statements. The projection equals the merged text byte for byte; the register is 474 rows, bijective with §31.

The regeneration

The renderer record moves (rust_renderer b2359a1c8…); the language-1.2 compiler-semantics ID moves to dcdf8a721708f0a9c5e57303f78eef5f5e37c58b7a11a04272d8670f618a699f (1.1 and 1.0 unchanged); the fixed point is reached at the first iteration; the production exports, the fixtures, the goldens and the calculus are rewritten or re-checked from this tree; the verification records of every project follow in the records commit, the Atlas's from the gate run's shard artifacts.

Identity cases run on the tree: rb_07, rp_10, rp_01, tc_03, bs_05, rb_15..rb_17, rb_19..rb_22, bs_06, bs_09, sp_22, sp_26..sp_30, sl_06..sl_08, ns_01, ns_05, rp_14, rp_18: all pass (bs_09 and sp_22 after a re-run alone; both run real Lean and failed only under the contention of 27 cases at once on the landing host).

Evidence

The gate on this PR; the Atlas join from its shard artifacts; CHANGELOG Unreleased carries the forward-merge bullet.


Generated by Claude Code

claude added 30 commits October 23, 2026 08:43
…ster statements, SPEC.md rows, and the failing named tests

The std runtime's types, impls, and helpers become a second module, RuntimeStd;
RB-20 states it, RB-19 states the associated function the validator learns for
it, BS-09 the two plants that follow (648 crates). The tests are red: the module
is not shipped, and the validator and its Lean twin refuse the associated call
only by name.
…ociated-call rule and its twin, the printer's compound assignment, the library table
…s records what a call returned and counted, the handwritten runtime and the printed one agree on them, and the golden transcripts are recorded (the last pass of the differential)
…ow, the SPEC.md row and paragraph, the scenario, the failing named test; RB-16, RB-17, RB-20 state the golden transcripts; the register order of RB-20 and two stale ranges
…ed, the cases compare the golden transcripts, and nothing holds runtime text but the printed modules
…--write, a changed-byte plant, and the statements say the comparison is against a recorded transcript of the removed oracle
…ranscript, its writer, and the red lines of the plants
…nts, the one transcript file named, the old counts, the module doc of runtime.rs, the unreachable decimal branch dropped, the decimal text's history and the largest module's Lean time
…nt for the number functions, RustStd with 17 more functions and theorems (unpruned definitions), the number text as methods of Str (statements and counts to finish)
…454)

The row, its scenario, its section-31 row and CONFORMANCE line, and the named
case conformance_sp_30, which fails on this head: the compiler table of runtime
items holds formatInt (the decimal text of an integer, called as format_i64),
and no theorem of RustRuntime or RustStd states it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ecimalInt

SP-29: the number functions are parse_int, format_int, and the format_ and parse_
function of each fixed width, one machine call each; formatDecimalInt is the
calculus primitive formatDecimal at the integers. RB-20: families 26 to 43;
format_int, which the handwritten runtime called as format_i64, is held to the
digest the transcript records for format_i64 on the same values; a format_int
that formats at another width is refused. The SPEC, CONFORMANCE and feature
copies of both statements are the ids.toml text byte for byte.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SP-29 lists refines_format_int; the RB-20 harness calls format_int on the values
of format_i64 and holds it to the recorded format_i64 digest (HELD_AS); SP-23
expects 44 machine constructors and BS-05 one more machine crate. Red until the
next commit adds the function, the primitive, and the theorem.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… refinement theorem

The runtime item FormatInt (the decimal text of an integer) had no function of
its own: the generated code called format_i64 for it, and no theorem stated it.
RustSyntax gains Machine.formatDecimalInt (family 43, one operand, no width),
whose meaning RustSemantics states as the calculus primitive formatDecimal, the
same as Item.formatInt; RustStd defines format_int as one call of it, printed
as Str::format_i64 because the printed runtime holds the integers as i64, with
the theorem refines_format_int; Item::FormatInt is called as format_int. The
bootstrap twin of RustValidate names family 43 and the item row format_int.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…5 counts

RB-20: a format_int that formats another value keeps its own label, which the
golden transcript lacks (HELD_AS), and the statement says so. The unit test
every_item_exists_with_its_failure_mode looks up the free function at the start
of a line: the number text has the method Str::parse_int, which the lookup took
for parse_int. BS-05: the four families 40 to 43 add one vector each to the
validate and correspond counts and to the evaluator counts (464/213, 327/100,
438/301), as the ten of H4b added ten.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eorems; the index plant without overflow

The shared term_cases list is pruned as H4b did: the 40 names H5 kept and the
eighteen number-function definitions (58) prove every theorem but the eight
format_<w> and parse_int; against format_i16 alone each of the 15 other names
was dropped (allWidth and hasWidth are needed), and against parse_int alone
(constructValue, evalArms and literal are needed); the 63 prove all 32.

RB-20: the plant "an index that finds the next position" computed i + 1 at
the harness index u64::MAX and panicked on the overflow without naming
index_list, so the case failed on the accepted H5 head 45f2f07 as here; with
wrapping_add the planted index returns the next item and the transcript names
index_list.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…SP-30, and the figures

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
claude and others added 24 commits October 26, 2026 04:59
…mit, not a second run of the gate on one runner

Rehearsals 3 and 4 showed the `just vv (release evidence)` job cannot
finish: the `bdd` recipe alone took over four hours on one hosted runner
and the job reached the six-hour limit before Atlas, so a tag would fail
there and `image`, `manifest`, and `publish` all wait for it. The CI
evidence §30.3 asks for already exists: `ci.yml` runs the whole gate as
the parts of one run on every push to the line. `.github/release-gate.sh`
finds the run that was the whole gate on the commit's tree (the commit,
or the second parent of a merge that lands its tree unchanged, as
`ci-plan.sh` reads a push), requires every job of it to have passed with
the Atlas join, every host, and every leg present, and writes
`vv-evidence.txt` with the run, its jobs, and the legs' retained
transcripts. SPEC §9.4 and §30.3 say so; `conformance_rp_18` requires
the gate job to run the script and no recipe, plants both, and runs the
script against planted runs.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…every conclusion refuses it

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…ry leg refuses it

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…t (RP-18, #34), with its plants

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…ap-h5

H5: the heap half's fifth landing on release/v0.4, with its Atlas records and the RB-21 path-separator fix.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…e branch

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
The release's evidence is the gate run of the tagged commit, not a second run of the gate on one runner.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…ixes that reached the line while the tag waited

The register is 457 and the three compiler-semantics IDs are the ones the
locks pin after H5; RB-21 ships; the std runtime types, impls and helpers
are in (RB-20), the number functions and the lowering stay 0.4.1.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…inement theorems; H5 printed their text

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
The 0.4.0 cut carries H5 and the release-workflow fixes; this is the v0.4.0 tag commit.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…mantics identity; locks, goldens, calculus, packages, fixtures and provenance regenerated

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
… reads

The first v0.4.0 tag run passed every job but `publish the release`,
which stopped at release-check: the §30.3 criterion reads the line
`just vv` prints when the whole gate passes, and the file the release
gate writes from the sharded run carried the run, its jobs and the
legs' transcripts, none of which says it. The rehearsals could not show
this, since `publish` runs on a tag only.

The script now ends the file with `gate: every job of run N passed on
<commit>` and `vv: the acceptance gate passed`, the criterion is one
function (`repo_model::release::ci_evidence`) with its own test, and
RP-18's planted-run test holds the script's output to that function.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…but the Atlas, on the moved identity

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
release-gate.sh ends the evidence with the line the release criterion reads; this is the v0.4.0 tag commit.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
The first pass's check-fixtures ran with an xtask the shared build
directory had taken from another worktree (its repository root is
fixed at compile time) and stopped when that worktree went away, so the
fixtures whose identity moved kept their old records; the first gate
refused vr-11, vr-15, cl-04 and cl-05 on their source ids.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…een shard artifacts of the first gate run

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…ap-h6

Stage 4 heap half H6 on release/v0.4: register 458, regeneration on the H6 identity, Atlas records joined.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
… release/v0.5

The generated files take the release/v0.4 side and are regenerated at the
landing; the hand-owned conflicts are resolved three-way (both RB-21 and
RB-22, the bootstrap validator counts as the sum of both sides, the
register count projected). The release/v0.4 specification text is ported
into the chapter modules of spec/ with cargo xtask spec port and
insert-after, and the register rows RB-21 and SP-30 join the Ids modules.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…rovenance on the merged identity

The renderer record moves to b2359a1c8 and the language-1.2 compiler-semantics ID to
dcdf8a721708f0a9c5e57303f78eef5f5e37c58b7a11a04272d8670f618a699f (1.1 and 1.0 unchanged);
the fixed point is reached at the first iteration, every project relocks, the production
exports and the fixtures are rewritten from this tree, and the model files are projected
(474 rows bijective with the register).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
…ut the Atlas, on the merged identity

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY
The plan and release-gate scripts run on the process PATH; the security
cases swap PATH for a git wrapper under the writer lock, so a decision
that ran under the swap failed closed to the whole gate. Both cases now
hold the reader lock for their decisions.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QppGKetUURseXowZun9KY

afflom commented Oct 27, 2026

Copy link
Copy Markdown
Member Author

First gate (run 40443244185): every job green but just vv (rest), where conformance_rp_14 failed in the planted-events check (the merge of an identical tree with a green run decided "whole gate" instead of "nothing"). Root cause, not a flake: plan_decisions and release_gate_decisions run their scripts on the process's PATH without the environment reader lock, while the SE cases swap PATH for a git wrapper under the writer lock (support::with_env); a decision that ran under the swap saw the wrapper's git and failed closed to the whole gate. The RP-14 failure of #298's arm job was the same race. Fix pushed as 83f9b26d0: both cases hold support::env_shared() for their decisions. The Atlas join from the first run's thirteen shard artifacts verified the Atlas (attestation 71ddcc506…) and changed no committed record, so the Atlas records stand as they are; the final gate runs on 83f9b26d0.


Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants