fix(cli): resolve a launched agent's flight on every orientation

[jlrickert]

Re-opens the work from #84. That PR merged into its base branch
(feat/mcp-bootstrap-mode) rather than main — #83 had been rebase-merged
seconds earlier, which rewrote the SHAs and left the base branch superseded but
undeleted. GitHub marked #84 merged, truthfully, into a branch that no longer
led anywhere. These two commits are rebased onto current main; nothing else
changed.

---

tap launch --agent NAME resolved the agent's flight once and exported the
result as TAP_FLIGHT. Environment variables outrank project and user config,
and a process cannot change its own environment, so the launched session was
pinned to whatever the flight was at launch. Editing the agent's flight and
calling orient again did nothing — and the session reported success while
staying on the old flight.

Export the reference instead of the value. TAP_AGENT names the agent, and
resolution reads agents[TAP_AGENT].flight out of the merged config on every
load. The variable is still constant — the agent driving a process genuinely
cannot change while it runs — but the mapping it points into is re-read at each
reload boundary, so a config edit plus orient now moves a running session.

Resolution happens once, at the end of ConfigService.load, rather than in
ActiveFlightName. load is what Reload re-runs, so late binding falls out
for free and every consumer — CLI, MCP orientation, keg listing, config explain — reads one already-resolved Flight() with no call-site changes. It
needs the env layer in isolation to tell a direct TAP_FLIGHT (which still
wins) from a flight that came from a file layer (which the agent overrides).

Precedence

1. --flight                      explicit runtime flag
2. TAP_FLIGHT                    direct env value (escape hatch)
3. agents[<active agent>].flight NEW - late-bound indirection
4. project  flight:
5. user     flight:              bootstrap baseline

Direct beats indirect, and naming an agent at launch beats an ambient project
default — which preserves what tap launch did before.

A TAP_AGENT naming an agent that is not configured warns and falls through
rather than failing: a session cannot fix its own environment, so a typo should
not brick a harness. The orientation payload names the active agent, so a reader
is told where the flight came from instead of being pointed at a flight: key
the agent silently outranks.

Codex's MCP registration gains TAP_AGENT in its env allowlist; TAP_FLIGHT
stays listed for a human overriding directly. Claude Code forwards its whole
environment and needs no change.

Telemetry

The reporter injects the active agent alongside the client version, so neither
call site has to remember to set it. This needs a matching hub-side field, which
is not yet released — until then an older hub rejects the batch, which degrades
to no telemetry rather than a retry loop thanks to the back-off fix in #81
(already on main).

Testing

• 8 cases over the precedence matrix, reload-picks-up-edit, and explain
  attribution
• 4 cases over a real config-driven MCP session, including
  edit-config-then-reorient-moves-the-session — the regression this exists for
• Verified end to end against the built binary

Both commits build, pass go test ./..., and are gofmt-clean individually.