Repository navigation
feat(tri): cache mutation rows so the full run can be finished (#2492) - #2533
Merged
Merged
Conversation
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
Contributor
This was referenced Aug 23, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Five operators over 21 gates passed twenty minutes and kept growing —
gft_backprop_microcode.pyalone has 47 sites, each a ten-second subprocess. The last two attempts were killed by timeouts, and one of those kills is what leaked two mutants into a branch (#2530).The cost had stopped being an inconvenience and started being a correctness problem: the full picture was the entire point of
--all, and nobody could reach it.Cached by what the answer depends on
The gate's bytes and the bytes of whatever control judges it. Both hashed, both must match.
The cache is written after every gate rather than at the end, so an interrupted run keeps what it measured and the next one resumes. Cold 2.6s → warm 0.5s on one gate.
Every reused row says
[cached]A cached green that read like a fresh one would be precisely the lie this command exists to find. The summary names the split: N measured, M reused.
I put that marker into one of two print paths, then two of three. The multi-column branch got it first; the single-operator branch printed cached rows identically to fresh ones; and the zero-site branch — "no failure path to break" — printed a third way with no marker at all. Three printers, one property, two corrections to reach all three. The same shape as
skip()in four copies, inside one function.Invalidation is verified, not assumed
On a planted repository: measure, reuse, then append one comment to the gate and watch the third run measure again. A cache that never invalidates is worse than no cache, and that direction is the one worth testing.
The stale case, stated rather than hidden
A fixture changing underneath a gate and its control leaves both hashes intact and the recorded row wrong. That is why the marker exists instead of silence: a reader who sees
[cached]knows which question to ask, and--freshanswers it.Refs #2492