Zero flake inputs: npins + haskell-flake standalone

[srid]

emanote's flake.nix no longer has any inputs. nixpkgs, haskell-flake, and the ema/lvar/heist-extra/unionmount/commonmark-*/emanote-template sources are pinned through npins, and haskell-flake is consumed via its standalone lib.evalHaskellProject API instead of a flake-parts module. Tree structure mirrors juspay/kolu — default.nix + shell.nix + small nix/*.nix files, so the same package graph drives flake and non-flake consumers.

Gone with it: flake-parts, nixos-unified, fourmolu-nix, git-hooks, the pre-commit plumbing that auto-generated .pre-commit-config.yaml, and the dormant omnix step hook. Formatting is now opt-in — just fmt runs nixpkgs-fmt and cabal-fmt; there is no forced hook, no fourmolu config, no pre-commit.

check-closure-size is preserved but reshaped: it's now checks.<system>.closure-size, a standard flake check that vira/nix flake check pick up via devour-flake — no external runner, no omnix. The legacy 600 MB target relied on justStaticExecutables + removeReferencesTo, both broken on current nixpkgs; until those work again, the budget sits at 8.5 GB to cover both systems (x86_64-linux 6.77 GB baseline, aarch64-darwin 7.53 GB) as a pure regression guard. The check uses pkgs.closureInfo + a summed du -sb — no hand-rolled graph parser.

Downstream users of inputs.emanote.flakeModule see no API change — the external flake-parts module (relocated to nix/flake-module/) is byte-equivalent, and internally delegates to the same mkSite function that the docs build uses, so a tweak to the static-site pipeline lands in one file not two.

Performance

Measured with hyperfine on x86_64-linux, 5 samples for nix develop, 3 for nix build (eval only), 1 for --rebuild. "Cold eval" means ~/.cache/nix/eval-cache-v6/*.sqlite wiped before each sample; fetcher/substitution caches stay warm (so deltas reflect the eval layer, not network).

Scenario	master (13 flake inputs)	this PR (zero inputs)	Δ
nix develop -c true — warm eval	170.0 ± 1.7 ms	162.3 ± 2.1 ms	−7.7 ms (−4.5%)
nix develop -c true — cold eval	2.905 s ± 81 ms	2.671 s ± 118 ms	−234 ms (−8.1%)
nix build .#default — warm eval	107.6 ± 0.9 ms	106.9 ± 0.1 ms	−0.7 ms (−0.6%)
nix build .#default — cold eval	1.560 s ± 12 ms	1.649 s ± 49 ms	+89 ms (+5.7%)
nix build --rebuild .#default — forced local rebuild of emanote	24.56 s	24.65 s	+0.1 s (noise)

Both branches produce the same emanote-1.6.0.0.drv derivation hash; rebuild cost is identical within noise.

Honest read: the perf case is weak. Warm eval is in the noise. Cold nix develop eval saves ~234 ms. Cold nix build eval is slightly slower. The gains are nowhere near the 2–3× that kolu saw on a much larger flake — because emanote's baseline is already modest (2.9 s cold vs kolu's 7.5 s). The real case for this PR is simplification: four dead dependencies gone, static-free default.nix/shell.nix entry points, closure-size guard now a first-class flake check. Merge for the cleanup, not the microseconds.

Try it locally

nix build github:srid/emanote/npins-haskell-flake-standalone#docs

nix develop github:srid/emanote/npins-haskell-flake-standalone

[srid]

Hickey/Lowy Analysis

#	Lens	Finding	Disposition
1	Hickey	Triple import ./default.nix in flake.nix	Fixed in this PR (7705932)
2	Hickey	Three npins load sites	No-op — templates belong at flake root; default.nix/shell.nix are legitimately independent entry points
3	Hickey / Lowy #1	Site-building logic duplicated across nix/emanote-site.nix and nix/flake-module/site/outputs.nix	Fixed in this PR (8455573)
4	Hickey	resolveLayer duplicates layer.nix outputs.layer/outputs.layerString	Fixed in this PR (a5822d8 — dead layer outputs removed after #3)
5	Hickey	shell.nix independent import	No-op — standalone nix-shell entry point is intentional (now shares project thunk with default.nix per elegance pass)
6	Hickey / pre-existing	Option duplication TODO between site module and home-manager module	Deferred — pre-existing scope
2	Lowy	nixpkgs version volatility well-encapsulated behind nix/nixpkgs.nix	No-op
3	Lowy	haskell-flake API volatility well-encapsulated in nix/haskell-project.nix	No-op
4	Lowy	Adding a new Haskell source: one line in packages + npins/sources.json	No-op
5	Lowy	Flake output shape: implicit contract between flake.nix and default.nix	Deferred — low magnitude, no schema yet
6	Lowy	Build tool scatter	No-op — collapsed by #1 fix
7	Lowy	emanote CLI invocation reimplemented in home-manager module + both site builders	Deferred — pre-existing, out of scope
8	Lowy	External flakeModule boundary isolated from internal consumers	No-op

Hickey rationale

Three separate import ./default.nix { inherit pkgs; } calls in packages/apps/checks meant the haskell-flake project was evaluated three times per system. Lifting them into a single perSystem thunk collapsed that to one. The other big one: nix/emanote-site.nix (internal, for the docs site) and nix/flake-module/site/outputs.nix (external, for downstream flake-parts users) both implemented the same static-site + live-server + link-check trio. Extracted mkSite as the single source of truth; the flake-parts module now delegates to it, which in turn made outputs.layer/outputs.layerString in layer.nix dead code (removed).

Lowy rationale

The volatility axes that matter are well-encapsulated: nixpkgs bumps touch one file, haskell-flake API changes touch one file, new Haskell deps touch one line in haskell-project.nix plus a pin. The one real leak the lens surfaced — "how a site is built" spanning two implementations — collapsed to one via the same refactor Hickey flagged. Deferred items (#5, #6, #7) are either pre-existing scope or low enough magnitude that a schema/abstraction isn't yet worth its keep.

Elegance follow-ups (police pass)

Four additional commits from the /code-police elegance pass:

• a2eb7887 — replace awk state machine with pkgs.closureInfo
• 78eaebcc — share haskell-flake eval across default.nix/shell.nix (saves one evalHaskellProject per system)
• 16a4ca52 — share mountPoint suffix across layer / layerString in resolveLayer
• 7c075e45 — trim WHAT-narration comments; keep only WHY

[srid]

/do results

Step	Status	Duration	Verification
sync	✓	1s	git fetch ok; forge=github; noGit=false
research	✓	15m 18s	Mapped current flake-parts layout, kolu target structure, haskell-flake standalone API, npins flow, and vira's devour-flake-based pipeline.
branch	✓	6s	On feature branch npins-haskell-flake-standalone.
implement	✓	6m 15s	npins set up with 9 pins, flake.nix rewritten with zero inputs, nix/{nixpkgs,overlay,haskell-project,emanote-site}.nix added, flake-module relocated, pre-commit/fourmolu/fourmolu-nix/git-hooks/flake-parts/nixos-unified removed.
check	✓	10s	cabal build all: up to date.
docs	✓	22s	nix/flake-module/README.md updated; user-facing docs already reference the external flake-parts module which still ships.
fmt	✓	9s	nixpkgs-fmt idempotent on all .nix files; cabal-fmt idempotent on emanote.cabal.
commit	✓	2m 33s	Commit 423024d pushed.
hickey+lowy	✓	13m 29s	Parallel review. Fixed H#1 (7705932), H#3/L#1 (8455573). Restored check-closure-size (a421000). Other findings deferred or no-op.
police	✓	7m 16s	Rules: no-dead-code (a5822d8). Fact-check: clean. Elegance: 4 commits (a2eb788, 78eaebc, 16a4ca5, 7c075e4).
test	✓	47s	cabal test all: 23 examples, 0 failures.
create-pr	✓	1m 3s	Draft PR #635 opened.
ci	✓	2m 26s	vira ci succeeded at HEAD 8a5cfde: 5 packages × 2 systems built, gh-signoff on both aarch64-darwin and x86_64-linux.
Total		52m 33s

Slowest step: research (15m 18s)

Optimization suggestions

• research was the dominant step (29% of total). The kolu structure mapping + haskell-flake standalone API research can be skipped on a re-run: both are now encoded in the tree. For follow-ups, pass --from implement so the workflow enters at the code-writing step.
• hickey+lowy (13m 29s) was the second-slowest — it had to run twice because the initial run was followed by a check-closure-size escalation from the user. If the closure-size scope had been part of the original prompt, the review wouldn't have needed re-anchoring. Lesson: surface critical preservation items in the initial brief.
• ci needed one retry because the 7.5 GB closure budget I set from the x86_64-linux baseline was 26 MB short for aarch64-darwin (7.53 GB). When a budget has to cover multiple systems, measure all targets before picking a number, not one.
• police spent 7m 16s across 3 passes + 4 elegance commits. A lighter pre-check during implement (a quick "would this pattern match pkgs.closureInfo?" sanity ping) could have landed the final shape in the first commit instead of requiring an elegance follow-up.

Workflow completed at 2026-04-21T00:26Z.