[CraigHutchinson]

Draft — exploratory groundwork, not ready to merge. Opened to stage the work for review; see the open items below.

Relates to #10.

What the recently-merged tag-matching fix (#9, de14713) did and didn't cover

It fixed tag filtering and made the untagged/no-tag case fall back to VERSION_FALLBACK instead of an empty header, and distinguished WARNING (tag matched but unparseable) from STATUS in the final summary line. It did not touch the mid-stream message(CHECK_FAIL ...) at the "Parse version" checkpoint, and did not touch template-file lookup at all. Both halves of #10 were still fully open going into this branch.

Part (a) — template lookup

CMake/Version.cmake:491 builds an exact-case path from VERSION_H_FILENAME; :496 does an exact-case existence check and silently falls through to auto-generating the wrong (C-only) template on a miss, with no error. The docs (:40 and Version.hpp.in:5) documented the lowercase "version.hpp" spelling against the shipped, capital-V Version.hpp.in — i.e. the docs themselves encouraged the exact casing that silently breaks. tests/test_cpp23_namespace/CMakeLists.txt:6 sets the lowercase spelling, so that in-repo test itself only passes today because it happens to run on a case-insensitive filesystem.

Reproduced with a genuinely case-sensitive filesystem, not simulated: enabled NTFS's per-directory case-sensitive attribute (fsutil file setCaseSensitiveInfo) on this Windows machine and confirmed two differently-cased files coexist there. Against unmodified master in that directory, with a consumer using the documented lowercase filename: template lookup missed, the wrong template was generated, and the consumer's main.cpp failed to compile with real MSVC errors (C2653, C2065) — this is a real, breaking consumer failure, not just an ugly log line.

Fix: case-insensitive fallback scan of the template directory when the exact-case match misses, with a CHECK_PASS note suggesting the exact case for portability; plus a new VERSION_H_TEMPLATE variable a consumer can set to an explicit path, bypassing the naming convention entirely (this reuses plumbing that already existed — it was already forwarded into the build-time cmake -P re-invocation — so it needed almost no new wiring). An explicit VERSION_H_TEMPLATE that doesn't exist is a hard FATAL_ERROR, not a silent fallback.

Re-ran the same case-sensitive-directory repro against the fix: log shows the case-insensitive match being found, and the consumer compiles, links, and runs, printing the expected version string. The explicit-override path was verified independently with a template at an arbitrary filename.

Part (b) — diagnostic wording

CMake/Version.cmake:402 emitted CHECK_FAIL for the ordinary "fresh clone / untagged branch" case, even though version_applyFallback (:453) then generates valid constants successfully a few lines later and the final summary (:468-471) correctly reports success. Reproduced directly on this (case-insensitive) machine — this half doesn't need a case-sensitive filesystem to demonstrate.

Fix: the "no tag reachable" case is now CHECK_PASS with wording stating the fallback is in use. A tag that matched but failed to parse — a genuine misconfiguration — correctly stays CHECK_FAIL. Doc comments in CMake/Version.cmake and CMake/Version.hpp.in updated to state the exact shipped filename and the new VERSION_H_TEMPLATE escape hatch.

Other verification

• Existing suite (ctest against tests/): all 24 tests still pass.
• No-.git/tarball scenario: a consumer tree placed entirely outside any git repo (confirmed via git rev-parse --show-toplevel failing), VERSION_SOURCE_DIR pointed at it — falls back to VERSION_FALLBACK correctly, final summary reads success, consumer compiles and runs printing the fallback version.
• FetchContent invocation mirroring the actual downstream CPM usage pattern (GIT_TAG <sha>, no manual include()) — configures, builds, and runs successfully against this branch's commit.

Still open before #10 can close

• A true Linux/ext4 CI run — everything case-sensitive above was validated via NTFS's per-directory case-sensitive mode on Windows, not a real Linux filesystem, per the issue's acceptance criteria.
• Decide whether to fix tests/test_cpp23_namespace/CMakeLists.txt:6 to the exact-case spelling (now optional, since the fallback handles it — but doing so keeps that test demonstrating the preferred spelling), or add a dedicated test that intentionally uses the wrong case to lock in the fallback behaviour long-term.
• A git-worktree-based consumer test case — not attempted.
• Check whether README.md also documents the old lowercase convention; not checked in this pass.
• General wording/style pass — this was written to prove the mechanism works, not polished for merge.

Local scratch artifacts from validating this (scratch/, build_test/, .scratch_casetest/) are untracked and not part of this PR.

[CraigHutchinson]

Superseded by #11, which merged into master while this branch was in progress and closed #10 directly.

Comparing the two: #11 is more complete than this prototype in two ways this session's groundwork pass didn't catch —

• It handles two differently-cased templates existing simultaneously (hard FATAL_ERROR on ambiguity); this branch's lookup didn't consider that case at all.
• It fixes a real bug in the old VERSION_H_TEMPLATE handling — the fallback path here still wrote into the package's own directory, which an installed or CPM-fetched copy doesn't own; Make header generation independent of filesystem case and git history #11's fallback writes to VERSION_OUT_DIR instead.

Real merge conflicts exist against current master in the same region of CMake/Version.cmake, and there's no remaining gap in #11 worth carrying this branch's changes forward for — including its own diagnostic-wording fix (issue #10, part b), which #11 already covers equivalently. Closing without merging.

For the record, in case it's useful corroborating evidence: this branch's validation independently reproduced the case-sensitivity defect via NTFS's per-directory case-sensitive mode on Windows (rather than Linux CI) and validated a FetchContent invocation mirroring the CPM GIT_TAG <sha> pattern a real downstream consumer uses — both against #11's fix, both passed.