docs(0110): close the synth split as won't-do — the tail is 29s

[adamkoot]

Close 0110 as won't-do — the synth tail is 29s, not 2–3 minutes

This PR started as a throwaway A/B measurement and is now the decision record it produced. ci.yml ships comment-only changes; everything else is lore.

Task 0110 assumed the Node/TypeScript tail on the rust job cost 2–3 minutes. Measured directly, it costs 29 seconds — and splitting synth into its own job would make CI slower. Closed under the task's own AC #2, which permits exactly that with the number recorded.

Method

Two commits pushed 95 seconds apart on this branch, so both runs shared runner class and cache state:

1. Baseline — comment-only change to ci.yml (run 30904595104, rust job 5m08s)
2. Experiment — the three tail steps deleted (run 30904701513, rust job 4m17s)

Both were reverted in 0aae4f4. Step-level timings, not job totals: Build Lambda bootstraps alone varied 17s between the two runs on identical content, so the 51s job delta is not the saving.

Result

	rust job	PR wall-clock
today	5m08s	5m08s
with a separate synth job	~4m39s	~5m29s – 5m49s

The tail is 29s, of which only ~20s actually disappears — synth's own 9s relocates into the new job. That job costs ~50–70s, serialized after rust because it needs its artifact, plus an unmeasured actions/upload-artifact step for ~110 MB.

The real cost centre is Build Lambda bootstraps at 3m24s — 67% of the job. This task optimized the wrong 10%.

What ships

• ci.yml — a comment block at the setup-node step carrying the number, so the next reader doesn't re-open this. git diff develop -- .github/workflows/ci.yml shows only # lines; every executable step is untouched and the 0077 guards are intact.
• Task 0110 archived as canceled / obsolete, converted to a directory.
• notes/G-option-1-synth-split.md — the complete, never-CI-tested implementation of the split, kept in case the arithmetic ever changes.
• lore/3-wiki/project/ci-pipeline.md — measured cost profile of the rust job.
• New task 0145 — cdk synth never runs on infra-only PRs.

Review

Findings 1–6 of the review on this PR are applied. The one that outlived the task: 0110 had ticked "synth still runs on every PR that could break it", and it does not — the rust job's paths filter omits infra/**, so an infra-only PR with a stack-level construct error goes green and fails at deploy. That is the 0070 failure mode, still open. The overclaim is withdrawn here and the gap is now 0145.

[karczuRF]

Code review — perf/0110_ci-split-synth-job-drop-node-from-rust @ e806236

Scope note: this PR is no longer what its title and body describe

The title says "MEASUREMENT ONLY — do not merge" and the body says "Either way this branch gets reset afterwards" and "The experiment commit deliberately removes a real guard (synth catches CannotFindAsset before deploy — task 0070). That is safe only because this PR is never merged."

As of e806236 that is no longer true: the head is two measurement commits, a revert, and a docs commit — a real deliverable that restores the guard and closes 0110 won't-do. A reviewer acting on the current body either blocks a merge that should land, or believes the head still has the guard removed. Please rewrite the title and body to describe the won't-do closure before merging. This review covers the merged effect of all four commits.

What checks out: the load-bearing measurements are accurate against the Actions API — runs 30904595104 (311s rust job; tail 7+13+9=29s) and 30904701513 (260s; Build Lambda bootstraps 204s→187s), SHAs match the two commits, pushed 95s apart. The ci.yml change is comment-only, parses as valid YAML, and keeps all 13 rust steps. Both changed files are prettier-clean under nx format:check --all. The revert is genuine: the blob at 0aae4f4 is 91e84b7, identical to develop.

---

1. status: done is not a valid status and fails the lore validator — medium

lore/1-tasks/archive/0110_PERF_ci-split-synth-job-drop-node-from-rust.md:5

lore/1-tasks/archive/CLAUDE.md enumerates completed / superseded / canceled; the repo uses completed 85 times vs done 4. Running lore-framework_validate against an existing status: done file errors:

Invalid enum value. Expected 'active' | 'blocked' | 'completed' | 'superseded' | 'canceled' | 'backlog', received 'done'

This file produces two such errors (frontmatter line 5, history line 28). Further: the archive convention for a won't-do closure is specifically status: canceled with a reason: key — which is exactly what this task is. The ad-hoc wont-do / decision tags on line 8 appear nowhere else in the repo and don't substitute for the status encoding.

2. The "byte-identical" claim is false in the shipped state, and AC #3 rests on it — medium

lore/1-tasks/archive/0110_PERF_ci-split-synth-job-drop-node-from-rust.md:38

The history note asserts "ci.yml is byte-identical to its pre-0110 state (blob 91e84b7), so the 0077 guards are untouched", and line 125 repeats it: "Byte identity with the pre-task blob is the proof."

But this PR ships ci.yml at blob f305ff6a, not 91e84b7 (git rev-parse e806236:.github/workflows/ci.yml). The doc contradicts itself at line 203, Design Decision 5: "the one deliberate deviation from byte-identity".

The change is a comment and is harmless — but the next reader who runs the stated verification gets a mismatch and has no way to tell whether the deviation is the sanctioned comment or a real edit. State the actual proof instead: "identical except an added comment block; git diff develop -- .github/workflows/ci.yml shows only # lines."

3. The documented recovery path for the option-1 implementation does not work — medium

lore/1-tasks/archive/0110_PERF_ci-split-synth-job-drop-node-from-rust.md:158

The doc says the ~82-line split implementation "is preserved in a git stash on this branch" and instructs:

git stash list | grep lore-0110

Git stashes are local to one working copy — they are not on a branch, not pushed, and not part of this PR. That command returns nothing in this clone (the only stash present is stash@{0}: On feat/0038_...: 0063-switch-cargo-lock).

Design Decision 4 (line 199) justifies stashing specifically so the code doesn't have to be rewritten if the arithmetic changes — but the only copy lives on one machine and dies with a git stash clear or a fresh clone. lore/3-wiki/project/ci-pipeline.md:100 repeats the claim. If the code is worth keeping, push it to a branch or paste it into the task file; otherwise drop both claims rather than sending a future reader after something that isn't there.

4. The synth-coverage claim (and AC #1) is wrong for infra-only PRs — medium

lore/3-wiki/project/ci-pipeline.md:45

The wiki presents make -C infra synth-production as "final proof the app the operator deploys actually synthesizes", and the archived task ticks AC #1 "Synth still runs on every PR that could break it."

But synth lives in the rust job, whose paths filter is packages/**, Cargo.toml, Cargo.lock, .github/workflows/ci.yml, tools/scripts/** — infra/** is only in the typescript filter.

Concrete scenario: a PR editing only infra/lib/*.ts that introduces a stack-level construct error runs the typescript job (nx run-many -t lint build typecheck plus verify-lambda-assets.sh), none of which invokes cdk synth — nx build is a tsc build, per infra/Makefile:4. CI goes green and the failure surfaces at deploy, which is the 0070 failure mode. verify-lambda-assets.sh only closes the asset-name-mapping subset of this.

Related: synth-cicd is never run by CI at all. The "Known gap" section records the develop-push hole but not this larger one, which is the gap most likely to bite.

5. The decision table double-counts synth's 9s, and omits the artifact upload — low

lore/1-tasks/archive/0110_PERF_ci-split-synth-job-drop-node-from-rust.md:113

Line 103 states synth relocates to the new job, and the new job's estimate at line 106 explicitly includes synth 9s. But the "after split" rust job is given as ~4m48s — only 20s below today's 5m08s. Synth's 9s is charged to the new job and left in the rust job. With synth relocated the rust job is ~4m39s, so the wall-clock and billed columns (~5m50s / ~5m48s) are each ~9s high.

Separately, the model omits the actions/upload-artifact step the split would add to the rust job for ~110 MB.

Neither error flips the conclusion, but these exact numbers are now quoted verbatim in .github/workflows/ci.yml:200-202 and lore/3-wiki/project/ci-pipeline.md:90, so the inconsistency propagates to three places.

6. The "billed minutes" column isn't a billing figure — low

lore/1-tasks/archive/0110_PERF_ci-split-synth-job-drop-node-from-rust.md:110

GitHub bills each job rounded up to the whole minute, at different per-minute rates per runner class. Today's single ARM job at 5m08s bills 6 ARM minutes; after a split it would be 5 ARM minutes plus 1–2 minutes on whatever runner the synth job used. Presenting 5m08s → 5m48s as billed minutes implies second-granularity, same-rate billing that doesn't exist.

The conclusion (the split is worse) survives under real billing too, so this is presentation rather than outcome — but a decision record whose entire value is the number should not put a number in the "billed minutes" column that no invoice would ever show.

---

Summary: findings 1–3 are the ones worth acting on before merge — the status value breaks the validator, and two documented verification paths (byte-identity, git stash) don't work for anyone who follows them. The measurement itself and the won't-do conclusion hold up.

[adamkoot]

@karczuRF ready for re-review at e54d147. All six findings applied plus the scope note.

#	Finding	Applied in
—	Title/body still said "MEASUREMENT ONLY — do not merge"	both rewritten to describe the won't-do closure
1	status: done fails the validator	canceled + reason: obsolete, frontmatter and both history entries
2	"byte-identical" false in the shipped state	replaced everywhere with git diff develop -- .github/workflows/ci.yml shows only # lines — currently 0 non-comment changed lines
3	option-1 impl unrecoverable from a local stash	committed as notes/G-option-1-synth-split.md; task + wiki repointed. This is what forced 0110 from a file into a directory
4	synth-coverage claim / AC #1 wrong for infra-only PRs	overclaim withdrawn in AC #1; wiki gained a Known gaps entry (incl. synth-cicd never running); spawned as 0145
5	table double-counts synth's 9s, omits the upload	corrected to ~4m39s rust / ~5m29s–5m49s wall-clock, double-count called out; the unmeasured ~110 MB upload-artifact noted as making the low end optimistic. Consistent across task, wiki and ci.yml
6	"billed minutes" isn't a billing figure	column dropped for an On billing paragraph. e54d147 also fixes the one place left stating it flat — the ci.yml comment now reads "saves nothing on billing", with 6 ARM minutes today vs 5 ARM plus 1–2 on the synth runner

Two things worth your call rather than mine:

• The validator still errors on this file — type: PERF (not in the enum; 6 tasks use it) and unquoted history dates (542 unquoted, 0 quoted repo-wide). Both are repo-wide schema drift, so fixing them here would leave 0110 as the only conforming file. Left alone deliberately; happy to spawn a task for the schema if you want it owned.
• The wont-do / decision tags you flagged are still on line 8. The status encoding is now doing the real work, so they are redundant rather than wrong — say the word and they go.

The measurement and the won't-do conclusion are unchanged.