[felix-andreas-finccam]

Summary

Brings loadfast into production shape: fixes real multi-package and lifecycle bugs found by probing, makes the package a clean installable artifact (R CMD check --as-cran: 0 errors, 0 warnings), and extends testing from 259 to 299 harness checks plus a new testthat suite that runs against the installed package during R CMD check.

Loader changes (R/loadfast.R)

Multi-package dependency tracking (the main gap): previously, after reloading a dependency, importer packages kept running against a stale snapshot and a plain load_fast() on them short-circuited with "No changes". Now:

• Reloading a package flags every cached package that (directly or transitively) imports it.
• A flagged package's next load_fast() auto-upgrades to a full reload, with a message explaining why.
• Full loads recursively reload flagged dependencies first, so a chain C -> B -> A converges with a single load_fast(C) after A is reloaded.
• No-change loads do not invalidate dependents.

Lifecycle fixes:

• Depends: packages are attached during full loads (clear error when missing) — previously ignored, so package code relying on them crashed.
• .onUnload runs best-effort when unloadNamespace() refuses (namespace imported by other loaded packages) before the registry entry is force-replaced.
• A failed full load no longer leaves a half-built namespace registered/attached; recovery on the next call still works.
• Renaming the package in DESCRIPTION in place now detaches/unloads the old identity instead of leaving it on the search path forever.
• Incremental re-sourcing follows Collate-aware full-load order instead of alphabetical basenames.
• Namespace spec version comes from DESCRIPTION, so getNamespaceVersion() reports correctly.
• Re-entrance guard is now a path stack: public load_fast() stays strict, internal dependency-reload recursion is allowed.

Packaging

• DESCRIPTION: version 1.0.0; methods/tools declared in Imports; dev deps moved from the non-standard DevDeps field to Suggests.
• roxygen2-generated NAMESPACE and man/ pages for both exports.
• Top-level message() moved to a suppressible .onAttach; sources are ASCII-only.
• library("testthat") replaced with a guarded attachNamespace(); tiny methods::: internals (.hasS4MetaData/.getGenerics/.TableMetaPrefix) inlined as equivalents so --as-cran is warning-free.
• Complete .Rbuildignore; renv settings track Suggests.
• R CMD check --as-cran: 0 errors, 0 warnings; the only inherent NOTE is the attach() call (same NOTE pkgload carries).

Testing

• Stage 6 (new) — multi-package dependency invalidation: direct importer invalidation, transitive chain convergence, no-change loads, whole-namespace import(), S4 class redefinition across packages, forced-unregister path incl. .onUnload.
• Stage 7 (new) — production hardening: Depends attachment + missing-Depends error, rename-in-place cleanup, failed-full-load cleanup and recovery, namespace version fidelity, Collate order on incremental reload, re-entrance guard, verbose logs.
• tests/testthat/ (new) — self-contained smoke/error-path tests that run against the installed package during R CMD check, so the check proves the built artifact works.
• Harness result: 299 passed, 0 failed (verified locally on R 4.3.3/Linux).

CI

• Harness job now also runs on macOS (Linux/Windows/macOS, R 4.5.2 via renv).
• New R CMD check --as-cran job on Linux/Windows/macOS at R release, plus oldrel-1 on Linux.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S8v6K52c7AJwuDkyyETMU7

---

Generated by Claude Code

[felix-andreas-finccam]

User test: built a fresh 3-package stack and ran it against main and this branch

I did an independent user test of this branch rather than a code read: I wrote my own toy
package stack from scratch, validated it against a real R CMD INSTALL + library() first so my
expectations match actual R semantics, then ran the identical scenario against main and this
branch in separate R processes.

Verdict: this is a clear improvement and I'd merge it. The multi-package tracking works exactly
as described, and the branch also fixes something bigger than the PR description gives it credit for
(see "The most important fix" below). A handful of small findings at the end.

Setup (click to expand)

• Ubuntu 24.04, R 4.3.3. CRAN was unreachable from my sandbox, so rlang, R6, digest,
  data.table, testthat and pkgload were built from their GitHub HEADs (rlang 1.3.0.9000,
  pkgload 1.5.3.9000).
• Both versions installed side by side into separate libraries (loadfast 0.0.1 from main,
  1.0.0 from a3f7ed7).
• My stack — three packages, deliberately covering every cross-package mechanism the README claims:
  • tsdata — S3 generic + methods, S4 class + generic + method, R6 class, an exported constant
  • tsstats — imports tsdata via importFrom / importClassesFrom / importMethodsFrom, plus tsdata:: calls
  • tsreport — whole-namespace import(tsstats), so tsreport -> tsstats -> tsdata is a transitive chain
• Sanity check first: R CMD INSTALL + library() on all three, confirming every assertion I make
  below is what real R does.

Scenario results

23 assertions over a realistic edit-reload loop. main 18/23, this branch 22/23.

Behaviour	main	This PR
Baseline cross-package access (importFrom, import(), ::, S3, S4, R6)	pass	pass
Direct importer picks up a dependency edit	stale — prints No changes in tsstats/R., returns stats-on-v1	Full reload of 'tsstats': dependency 'tsdata' was reloaded. → stats-on-v2
Transitive importer (tsreport) picks up the edit	stale	correct
S4 method redefinition visible in importer	stale (3 kg, should be 3 kg!)	correct
getNamespaceVersion()	0.0.0	0.1.0 (matches DESCRIPTION)
Failed full load leaves the namespace registered?	yes — tsbroken stays in loadedNamespaces()	no, cleaned up
Rename Package: in place	both old and new on search()	old detached, new attached
Depends: on a missing package	silently ignored	clear error naming the package
Collate-order on incremental reload	alphabetical (wrong value: 30)	Collate order (correct: 150)

One nuance worth recording, because it explains why main doesn't fail every cross-package test:
staleness on main depends on how the importer reaches the dependency. tsdata::series_scale(...)
resolves live through the registry, so an edit is picked up even without reloading the importer;
an importFrom-imported value like DATA_LAYER_VERSION is a snapshot and goes stale. So on main
the failure is intermittent by access style, which is the worst kind. This PR removes that.

Claims I verified independently

• "No-change loads do not invalidate dependents" — confirmed. load_fast(tsdata) with no edits leaves tsstats on the cheap No changes path.
• "A chain C -> B -> A converges with a single load_fast(C)" — confirmed. After editing and reloading tsdata, one load_fast(tsreport) emits Reloading dependency 'tsstats' first: ... and lands on the right value.
• The flag is one-shot — first call after a dep edit is a forced full reload; the second and third are back to No changes. No sticky full-reload penalty.
• Cascade is conservative — I broke tsdata on disk after flagging its dependents, then reloaded from the top. The cascade rebuilds tsstats/tsreport against the loaded tsdata namespace and never re-sources the package I didn't ask it to reload. Good call.
• Lifecycle hooks — .onUnload then .onLoad fire correctly for the importer during a forced full reload.
• load_fast_register_reload() still works after a forced full reload.
• Test suites — harness: 299 passed, 0 failed on this branch vs 259 passed, 0 failed on main. The new tests/testthat/ suite runs under R CMD check: [ FAIL 0 | WARN 0 | SKIP 0 | PASS 24 ].
• R CMD check --as-cran — the only inherent NOTE is the attach() call, as stated. (My run also showed notes/warnings for missing pandoc, no LaTeX, no CRAN index and an unavailable locale — all sandbox artifacts, not the package.)
• Base-R-only runtime — I installed the built package into a library containing only loadfast and R6, with no rlang anywhere on .libPaths(). The full stack loads and every cross-package feature works.

The most important fix here isn't in the PR description

main calls rlang::ns_registry_env(), which rlang gates on the R version:

# rlang/R/env-special.R
ns_registry_env <- function() {
  if (getRversion() >= "4.6") {
    deprecate_stop("`ns_registry_env()` is defunct.")
  }
  .Call(ffi_ns_registry_env)
}

So on R >= 4.6 with rlang >= 1.2.0, main's load_fast() fails on every full load.

I verified this rather than inferring it: I rebuilt rlang with only that version gate flipped
(>= "4.6" -> >= "4.3"), so the function takes exactly the branch it would take on a real R 4.6,
and ran main against it:

ns_registry_env():  STOPS -> `ns_registry_env()` is defunct.
load_fast():        ERROR -> `ns_registry_env()` is defunct.

The same run on this branch works, because commit e973af6 replaced it with runtime .Internal()
calls. That's a "loadfast stops working entirely on the next R release" fix, currently buried in the
second commit's message. I'd promote it to the top of the PR description and add a NEWS entry —
it's the strongest reason to merge, and users on main will want to know.

What the new correctness costs

Synthetic package: 401 R files, 10,002 functions, 202 exports (the realistic shape — exporting all
10k made parseNamespaceFile superlinear and skewed my first measurement).

Operation	main	This PR	pkgload::load_all
Cold full load	0.367s	0.347s	1.00s
Reload, one own file changed	0.012s	0.012s	0.85s
Reload after a dependency edit	0.007s — stale	0.317s — correct	0.80s — correct

So the honest tradeoff: a leaf reload after touching a base package goes from ~7ms to ~317ms,
roughly a cold load and ~26x the incremental path. That is the right default — a 7ms wrong answer
is worth nothing — and it is still ~2.7x faster than load_all on the same package, with the common
case (editing the package you're in) untouched at 0.012s. Worth one line in the README so nobody is
surprised by the cliff.

---

Findings

1. NAMESPACE changes are still not detected (pre-existing; affects both versions).

Adding a function and its export() and reloading:

| Incremental reload: 1 changed [aaa_constants.R]
-> tsdata::series_range : ERROR: 'series_range' is not an exported object from 'namespace:tsdata'

full = TRUE fixes it. This isn't a regression, but it lands squarely on this PR's audience —
adding an export to a base package is routine multi-package work, and the README's full = TRUE
guidance only mentions deleted files and removed functions, so there's nothing pointing users at it.

Suggested fix: hash NAMESPACE (and DESCRIPTION) alongside the R/ files and force a full load
when either changes. One extra md5sum per call, it makes the existing rename-in-place detection
fall out naturally, and it closes the gap rather than documenting it.

2. The README overstates the new guarantee.

a plain load_fast() on an importer never silently reports "No changes" while it is running
against a stale dependency snapshot.

It does, if the dependency was edited but not itself reloaded — which is the natural thing to do
when your cursor is already in the importer:

(tsdata edited on disk, load_fast(tsdata) NOT called)
| No changes in tsstats/R.
-> tsstats::stats_stack_version(): stats-on-vX      # correct value: stats-on-vTRAP

The very next paragraph states the correct workflow ("reload the package you edited, then reload
whatever package you are working in"), so this is just a wording fix — something like
"...never reports No changes after you have reloaded a dependency it imports from". Worth
getting right, since it's the sentence a user will rely on.

(Genuinely fixing it looks cheap, if you want it: the cache already holds every loaded package's
path and per-file MD5s, so a forced load could re-hash cached dependencies and flag them. Happy to
leave that for a follow-up.)

3. The comment at R/loadfast.R:6-7 is imprecise about the trigger.

rlang's ns_registry_env() is defunct as of rlang 1.2.0

It reads as unconditional, but it's defunct in R >= 4.6 — rlang 1.2.0 is where the version gate
was added; on older R it still works (which is why main still passes its suite on R 4.3/4.5).
Suggest: "defunct in R >= 4.6 (rlang >= 1.2.0)". Given this is the load-bearing justification for
the .Internal() approach, it's worth being exact.

4. The import-failure error could carry the same hint the new Depends error does.

The new Depends error is genuinely helpful:

Failed to attach Depends package 'foo': there is no package called 'foo'
If it is a source package, load it with load_fast() first.

The import path has no equivalent, and it's the error a multi-package user actually hits first:

Import failed for "cyca": there is no package called 'cyca'

Adding the same second line there would help exactly the audience this PR targets. (Reached via a
cyclic import() in my test, but any not-yet-loaded source dependency produces it.)

Related and worth noting as an improvement: on main that cycle appeared to load, because the
failed first load left a half-built namespace registered for the second one to import from. This
branch's failed-load cleanup means both packages now fail cleanly instead. Correct, if initially
surprising.

---

None of these block merging — 1 and 2 are the two I'd want addressed, and 2 is a one-line doc edit.

---

Generated by Claude Code