[Sarbajit-Sengupta]

Add eight tested improvements to TraceCache:

• Reject input/output hard-link aliases.
• Atomically replace individual output files and preserve prior output on interrupted writes.
• Stream gzip input with decompressed-size limits and corruption diagnostics.
• Import strict CSV keys, including quoted multiline values.
• Carry LRU state across warmed phase windows, retaining independent cold-start mode.
• Avoid unnecessary FIFO/Belady replays at known capacity boundaries.
• Verify saved schema-2 reports by recomputing all declared fields and the trace fingerprint.
• Benchmark phased, uniform, cyclic and single-hot-object workloads with baseline parity checks.

Validation: 88 local tests pass, branch-aware coverage is 99.24%; Ruff, formatting and strict mypy pass. Wheel/source builds and clean-wheel CLI analysis/verification pass. Four benchmark families retain five shuffled timing repetitions and separate allocation peaks; direct replay wins for one capacity in all four families, and the single-hot-object workload shows no meaningful profiling speedup even at 64 capacities.

Scope: individual output replacement is atomic; JSON+SVG is not a transaction and power-loss durability is not guaranteed. Gzip fingerprints cover decompressed bytes. Report verification checks consistency, not collector provenance or authenticity. Schema-1 reports require regeneration. No package registry release, deployment, production adoption or independent human review is claimed.