[ToikY]

The current full-history checkout can omit historical blobs with filter: blob:none, but replay still needs exact declared sources. This prototype shows that the unchanged package validator automatically hydrates a genuinely missing old/deleted source on a controlled local partial fetch. A separate client rejects the same source while offline, then passes metadata, publication-history and mode checks after explicit hydration.

Only research/ci-blob-filter-compatibility is added. Actual workflows remain unchanged; the proposed two input lines are inert PROPOSED-NOT-APPLIED.patch data. This is compatibility evidence for the current checkout/source-reader seam, with production adoption and hosted performance left open.

Evidence:

• Independent fixtures agree on all 17 strict semantic outcomes: full refs/commits/trees/current HEAD bytes and modes, automatic hydration, explicit offline gates, and intended shallow/foreign-source/byte/mode/staged/unstaged rejection. Common projection SHA256: f46ad200b4f6c1edfd48e29281340a55a8764c511876ffb58d1dc60ca8b945f1.
• Four registered normal/optimized commands pass. The corrected 36-file export passes integrity checks and all four direct cold commands from read-only files, without a Git checkout and with Git-only PATH. External Git remains required.
• Twenty-eight actual own/peer receipt-schema mutants reject for intended reasons. Final committed bytes, executable modes and declared source history pass at local commit 50ccee4807762d32138ce29c81114cbc8d1b5974.

VALIDATION.md, RUN-RECORDS.json and retained first artifacts distinguish the original failed launcher from the first complete run and later review revisions. They also preserve the corrected timestamp, comparison-newline instrumentation error, and failed CRLF-normalized commit/export. A package-local attribute plus reindexing now preserves the official source bytes. Source/case inputs and the 17 outcomes remain unchanged through those corrections; raw temporary diagnostics are outside semantic equality.

Reproduce with CPython 3.12 and Git on POSIX:

python3.12 -B scripts/research_package.py --package research/ci-blob-filter-compatibility
python3.12 -B scripts/check_research_commit.py --package research/ci-blob-filter-compatibility --check-source-history --check-modes

Local evidence uses managed CPython 3.12.14 and Apple Git 2.50.1, file-only transport, fixtures below 50 MiB and bounded commands. It establishes neither hosted speedup nor immediate post-checkout offline operation, private-repository authentication, hidden dependency closure or arbitrary-command equivalence. Hosted CI did not certify this result: originalrun41719745758 planjob126209056703 cancelled during actions/checkout at its10minute limit,21:13:00–21:23:02UTC; setup/helper/selection/replay skipped. No rerun or dummycommit. No κ, frontier, merge or live-workflow change is proposed; maintainer edits and auto-merge remain disabled.

Credit: Douglas Colkitt and retained tooling contributors; #3304 publication ancestry, #3142/#3102 sparse discovery, #2505 complete batching/concurrency, #1595 batched source reading, #4984 maintenance-free fixtures, and official Git/actions/checkout. Exact predecessor pins, original headers and MIT license are retained in SOURCES.md/NOTICE. This adds the source-specific compatibility/hydration experiment, not those underlying mechanisms.

Public head: aba156a1131a99ef054fd2f4407ac5718a4959c0.