feat: bggrep tool + bgtail delta tailing

[lloydsk]

Stacked on #1 (project-local jobs dir) — review/merge that first; this branch contains its commits plus the new work below.

What

Two context-efficiency features for reading job logs:

bggrep — capped in-extension pattern search

• Regex search over a job's log: line-numbered matches (grep -n style), optional context lines with …[N lines skipped]… gap markers, capped at 50 matches + the ~8KB condenser budget.
• Runs inside the extension with native fs access, so it works on any jobs dir — including global logs that project-sandboxed tools (ctx_execute_file) cannot reach. That gap repeatedly forced agents into uncapped raw greps; bggrep closes it.
• With no pattern, a generic failure-signature default is used — deliberately a convenience only (overridden by passing your own pattern): a general-audience tool across languages/task types misses more than it catches, so heuristics are never load-bearing.
• Record-first log resolution (correct across mid-session config changes), same as bgtail.

bgtail delta tailing

• First read for a job: full last-N tail (unchanged semantics).
• Repeat reads: only lines appended since your last read, with a +N new lines header — polling a running job never re-pays context for lines already seen; a no-change poll returns a single tiny line.
• Bookmark = high-water mark (total content lines + byte size at read time); deliberately-skipped prefix lines are never replayed as "new"; a shrunken/replaced log (detected by line count or bytes) resets to a full tail with a note; raw: true keeps the verbatim last-N window but still advances the bookmark; bookmarks are in-memory only — a session restart starts fresh.
• The /bgtail slash command inherits everything via the shared core.

Testing

tsc clean; 51/51 node:test — 8 new: bggrep (line numbers + explicit/default/no-match, context windows + gap markers, invalid-pattern error, 50-match cap, record preference) and delta tailing (full→new→none sequence, raw-advances-bookmark, byte-level shrink reset). The delta rewrite also fixed a regression caught by existing tests: the exit marker is now filtered before the window is sliced, so "last N lines" keeps meaning the last N content lines.

Docs

README tools table + two-tier read model; SKILL.md tools table + reading guidance.

[lloydsk]

Addressed the fresh-eyes review in f7808cf (56/56 tests, tsc clean):

Finding	Resolution
Replacement log with same line count/byte size slips past shrink check	Fixed — bookmark now stores the first content line; append-only logs never mutate line 0, so a changed first line = replacement → full-tail reset. Invariant-based, catches rotation regardless of size
bggrep empty log: "in 1 lines" via "".split("\n") === [""]	Fixed — in 0 lines
CRLF: trailing \r breaks $-anchored patterns, leaks into output	Fixed in both bgtail and bggrep (/\r?\n/ split)
No schema minimums: lines: 0 → slice(-0) returns everything; negative context drops the match lines themselves	Fixed — schema minimums + defensive clamps in the shared cores (slash-command path bypasses schemas)
Dead branch in bgtail's empty-body ternary	Removed (delta early-return makes it unreachable)
2KB/line cap undocumented	Fixed in README + SKILL.md wherever ~8KB appears
Test gaps	Added — replacement reset, CRLF, empty+notFound, context+cap interaction, param clamps
Relative jobsDir traversal → useless ../ exclude pattern	Hardened in this branch — appendExcludePattern skips ../-prefixed patterns (unreachable via the walk-up today; defense-in-depth)

Declined by design (from the review's request-changes on #1): ancestor walk-up for subdirectory cwds — an adjudicated design decision; the README's Rules section documents the actual behavior (cwd must itself be a project root, else global fallback). Path traversal in jobsDir is user-owned config, not an attack surface.