test(self-packaging #46): integration tests for round-trip behaviours

[jrosskopf]

Part of epic #40. Stacked on #55 -> #54 -> #53 -> #52 -> #51.

Summary

End-to-end integration tests against the real flapi binary,
covering the spike's documented round-trip behaviours. Plus an
implementation fix surfaced by the test for the secret deny list.

• New file: test/integration/test_self_packaging.py -- 8 tests,
  pytest, runs in ~24 s on Linux x86_64 debug build.
• Fix to src/pack.cpp -- validate input tree before copying the
  multi-GB host binary, so a rejected .env doesn't leave a
  half-written bundle on disk.

Coverage map vs. the spike's 9 behaviours

Spike #	Behaviour	This PR
1	filesystem mode serves endpoints	deferred -- existing tavern suite already covers filesystem-mode HTTP
2	flapi pack produces a bundled binary	✅ test_pack_produces_bundled_executable
3	bundled binary serves endpoints from clean cwd	deferred -- needs bundled-mode config-path resolution work
4	flapi info reports EOCD offset + entries	✅ test_info_reports_bundle_contents
5	self-hosting (bundled binary packs new bundled binary)	✅ test_self_hosting_rebundle
6	re-pack idempotence	✅ test_repack_idempotence (3 rounds)
7	1-KiB-truncated bundle falls back to filesystem mode	✅ test_truncated_bundle_falls_back_to_filesystem
8	flapi unpack dumps the archive	✅ test_unpack_restores_all_entries
9	read_csv('embed://...') in an endpoint template	deferred -- in-process proof-of-life already in #54's unit tests

Plus two non-spike invariants:

• test_pack_refuses_secret_files_by_default -- .env in input
  causes non-zero exit AND no half-written output binary.
• test_pack_with_allow_secrets_bundles_env_files -- --allow-secrets
  bypasses the deny list.

$ FLAPI_BUILD_TYPE=debug uv run pytest test_self_packaging.py -v
============================== 8 passed in 24.38s ==============================

Why three behaviours are deferred

Behaviours #1, #3, #9 all need server-lifecycle test infrastructure
(subprocess start, health-wait, hit HTTP endpoints, teardown). #3 and
#9 additionally need the bundled binary to load its embedded config
when run from a clean cwd -- which depends on ConfigManager's
template-path resolution behaving sensibly when the loaded config
came from embed://flapi.yaml. That's a non-trivial integration
question worth its own scope.

The DuckDB read_csv('embed://...') proof-of-life already runs
end-to-end in #54's unit tests against the in-process DuckDB
instance, so #9 is verified at the layer that matters most.
Filesystem-mode HTTP is exercised by every existing tavern test.

I'll file the server-lifecycle integration tests as a follow-up
issue once #51-#55 land.

Implementation fix in src/pack.cpp

Before this PR, Pack() did:

1. copy 1.7 GB host bytes to <out>
2. walk input dir
3. throw PackError on .env
4. leave a 1.7 GB half-written file in the user's tmp dir

Re-ordered to:

1. walk input dir (validates against deny list)
2. write archive in memory
3. copy host bytes
4. append

Now .env rejection is fail-fast, no garbage left behind. The
test_pack_refuses_secret_files_by_default test enforces it.

Test plan

• All 8 tests pass locally (Linux x86_64, debug build).
• Existing C++ unit tests still pass (no logic change beyond
  Pack() ordering, which is exercised by test_pack_idempotence
  and test_pack_refuses_secret_files_by_default).
• CI cross-platform once the whole stack lands (test(self-packaging #46): integration tests for round-trip behaviours #56 covers that).

Closes #46 (partial -- server-lifecycle behaviours deferred to a
follow-up).