Fix native CI disk exhaustion and lease handoff flakiness

[SaladDay]

Problem

The native Harness job fills the GitHub runner disk while compiling the debug test graph. It also shares disk with the subsequent release build. Separately, the API suite intermittently starts a successor immediately after a client connection closes, before PostgreSQL has released that owner's advisory lock.

Changes

• Run native checks and the unmodified release build on separate runners, disable incremental compilation, cache only downloaded dependencies, reclaim unused Android/.NET SDK space and report disk usage.
• Synchronize the affected test handoffs on the old owner's exact database lock disappearing, with bounded waits and immediate query-error failures. Production lease behavior and startup rejection are unchanged.
• Document both CI resource policy and the test synchronization boundary.

Validation

• make check passed on Linux with a dedicated PostgreSQL 16 database.
• The two affected test functions each passed 100 consecutive runs.
• Native packaging checks, actionlint and git diff --check passed.
• GitHub core-check, agents-api, actionlint and native Harness checks passed. Native checks retained 23 GB free disk with 5.7 GB in the target directory; the independent release build reached the original 60-minute timeout with 21 GB free disk; no compiler error was reported.
• A fresh independent-context blind review of d115653 against 2862fce found no in-scope blockers. The reviewer checked all changes and confirmed the native gates and production ownership behavior are preserved.

No tests, Clippy checks or release targets are removed.

Private evidence is retained at ~/.parsar/tests/parsar-core-ci-20260921/ on the validation host and ~/.parsar/parsar-core-ci/ on the development host.

• make-check.log: SHA-256 d068d1da5c01a5653c18d50eacd76ec19d10a1b0dd47aeb6650f594be0c006fb
• repeated-tests.log: SHA-256 b58cab4b8bdceb4f7a51db577708e379b1da91160474ede9df994fec5a21eac9

Final CI adjustment and merge scope

Commit 44b553f gives only the cold optimized release build a 120-minute limit; checks retain 60 minutes. Actionlint and whitespace checks passed after this workflow-only adjustment. The extended release run is not yet verified. The user explicitly requested merging and letting the main-branch workflows run without continued monitoring. Earlier check results apply to d115653; no runtime or test code changed in the final timeout adjustment.

A second fresh independent-context blind review of the full final diff (44b553f against 2862fce) found no evidence-based blockers.