[kacy]

Four leak fixes, one commit each. They share one PR because each one regenerates the bootstrap seed, so separate PRs would conflict on it.

• a discarded catch, select or match releases only a string result #1091: a match or select thrown away at statement position left its arm value in the temp with nothing to release it. The checker also never recorded the type of a statement match, so the emitter could not tell what the temp held. The discard now releases the temp for every heap kind when every arm is owned or retained on the way in.
• leak: a fresh String result passed straight to assert_eq is never released #1118: this was already fixed on main when assertions began releasing their owned arguments (feat: count checker covers 97% of corpus functions, and the leaks it found #1361). The commit adds a leak case so it stays fixed.
• a match on a result local with a tuple-shaped payload releases only the box #1150: the cascade scan refused every match on a result local, so the local fell back to a plain struct release that freed the box and not its payload. A bare arm binding is the box itself with no count, so the local now cascades when the arms only read it.
• a result box held by a struct field or a list element is never reclaimed with its payload #1167: a result box owns its payload but has no destructor, so a struct field or container element that held one freed the box and stranded the payload. A result-typed struct field was not recorded at all, so the box leaked too. A box stored into a holder now takes the payload destructor. Under the box abi a function that returns a shared box now hands back its own copy (__result_unshare_*), because every caller that opens a call's box takes the payload's count over. This also fixes an invalid read the box abi had on fwd(rs)! over a list element.

Each fix has a leak case under tests/leaks and a regression case with expected output in the memcheck set. The count checker already flags the discarded-match leak once a program has it, and the new corpus case puts it there. It learned the unshare helper's ownership. It cannot see the #1150 and #1167 leaks, because a plain struct release of a result box looks balanced unless something read the box's ok or err slot, and the marks do not say a box carries a counted payload. Modeling that needs a new mark and is left for later.

Closes #1091
Closes #1118
Closes #1150
Closes #1167