fix: harden skill discovery and preserve loaded instructions

[MelodyVAR]

Skill loading had inconsistent behavior across discovery, invocation, and compaction: directory symlink cycles caused repeated traversal, --no-skills still scanned skill trees, nested ignore rules lost their scope, and multi-line /skill:name commands were not expanded. Step also advertised read when its active tool was read_file.

Fix discovery with canonical directory visitation and a shared directory-scoped ignore matcher. Skip automatic skill collection before traversal when disabled, while retaining explicit CLI paths and correctly identifying skills-only packages. Align the SDK helper with project-first precedence and accept BOM-prefixed skills in the execution-environment loader. Generate the catalog instruction for the active read tool and split skill commands on whitespace.

Built-in compaction now preserves actually loaded skill instructions independently of the generated summary. Successful reads and explicit invocations are recorded with their locations/read ranges, carried in compaction metadata, and appended verbatim to the resulting context. Full rereads replace older content, repeated ranges are deduplicated, and reused tool-call IDs are correlated in message order. Step's read/write/edit aliases also participate in file tracking. Custom compaction extensions retain ownership of their summaries.

Validation:

• coding-agent full suite: 3514 passed, 55 skipped.
• agent-core full suite: 465 passed, 1 skipped.
• CLI full suite: 624 passed.
• 27 focused regression cases cover discovery, invocation, encoding, precedence, and compaction, included in the counts above.
• pnpm run build:offline passed.
• All checks from pnpm run check passed, with Biome run without --write.
• Isolated two-self-symlink probes complete for the low-level loader, the full resource loader, and --no-skills; the full loader completes in about 50ms where the original exceeded a 3s external scan limit.