Copying ::punk::libunknown::epoch verbatim into a fresh interp/thread
carries per-epoch 'added' maps that drive zipfs_tclPkgUnknown's
can_skip_sourcing short-circuit - sound only in the interp that performed
the scans. Packages the donor obtained at an earlier epoch (or with no
unknown call: Thread is baked into thread::create'd threads) are absent
from the donor's current-epoch map, so recipients skip their pkgIndex
sourcing and requires fail where stock tclPkgUnknown would succeed.
Bit deterministically as repl_console_backend_tkconsole under the suite
9.0.5 runtime: 'package require Tk' bumps the epoch pre-init, the
codethread inherits the post-bump sparse map, the code interp then can't
find vfs/Thread/registry, its init aborts before the G-001 conredirect
junction vars exist, and the first pending-collect error left the repl
reader inactive until the driver failsafe. Tcl903 escaped only by an
accidental extra epoch bump in its code interp (fresh map, full rescan).
Fix: punk::lib::libunknown_epoch_export (PUNKARGS-documented) - epoch
record with every 'added' map emptied; counters/untracked/scripts/index
globs preserved. Used at all four propagation sites (interp_sync,
snapshot, repl %lib_epoch%, repl safe-arm). Hardening: the repl
pending-collect now tolerates missing junction vars so a code-interp init
failure can't wedge the reader.
Also: test socketpairs pinned to -myaddr 127.0.0.1 (all-interfaces
listens raised windows firewall consent popups for each newly built test
exe - user datapoint).
Verification: consolebackends 3/3 both runtimes (suite 42s->14s); full
runtests sweeps identical on suite runtime and Tcl903: 1037/1021/15/1
with the sole failure exec-14.3 (documented machine baseline). G-098
acceptance sweep met. Pre-existing PUNK_TEST_TK opt-in quirk noted in the
goal (silent early end, both runtimes, out of scope).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Julian Noble
d3cc914a52